ark · manifest of references for an offline AI mirror · generated 2026-10-17

What ark references

The full manifest as a drill-down: source repositories by purpose, model weights by kind, datasets, reference data, assets, containers and packages. Every entry links to its upstream. This page is generated from manifest/*.tsv by report/gen-ark-directory.py; nothing here is a local path.

730source repos · 35 categories

190weight repos

29datasets

146reference sets

14container images

154extra PyPI packages

Source repositories730 · 35 categories

Latest-only shallow clones. Categories are by purpose; add a line to manifest/repos.tsv to contribute.

Inference engines44 repos

0xShug0/audio.cpp

note ggml speech backend: ASR/TTS/diarization GGUFs (the cstr/* ladder aimbot indexes)

antirez/ds4

AtomicBot-ai/atomic-llama-cpp-turboquant

note llama.cpp fork: TurboQuant WHT-rotated KV cache/weight compression, Gemma 4 MTP + Qwen 3.6 NextN spec decode; base of halo's NPU backend (patches in halo experiments/atomic-llama/patches)

note Gufo: Strix Halo (gfx1151) inference engine, one host: Qwen3.8 27B/Flash-Next, DeepSeek V4 Flash, Qwen3 ASR/TTS, image/video; aimbot backend; nix or podman toolbox build

halo-box/llama.cpp

note Halo Box near-mainline llama.cpp fork: --spec-prefill, --spec-draft-adaptive, RDNA 3.5 Vulkan fixes, hidden server presets

halo-box/strix-llama.cpp

note llama.cpp fork tuned only for Strix Halo (gfx1151): RDNA 3.5 Vulkan mat-vec chunking (spec decode), ROCm MMQ/MoE/GDN/FA kernels, ROCmFPx quants, multi-point reasoning budget; superset of halo-box/llama.cpp; has upstream RPC RDMA incl. completion channel (#29440)

halo-box/strix-llama.cpp#62e9d99c4854212986e62786ab5d3eb07d284fdc

note strix-llama.cpp PR #172 (open 2026-10-12): HIP RDNA3.5 flash attention + fused GDN/MoE decode for Qwen3.5/3.6/3.8 on gfx1151; candidate for POD-LLAMA rebuild (qwen38-27b on pod-control)

halo-box/strix-llama.cpp#79194d304b3812621a94b1f0f81852cd4ccc02c7

note strix-llama.cpp PR #56: HIP RDNA3.5 MMVQ/MMQ kernels + per-type dispatch thresholds for small-batch quantised matmul (decode, spec verify)

halo-box/strix-llama.cpp#94f8b99864a8ea441741ad5e1ddb561b02db9118

note strix-llama.cpp PR #108: MTP draft-head vocab subset + even drafts / small-batch verify for qwen4exp

halo-box/strix-llama.cpp#9fbe5308fabbc79a53ddee5eadfad74ae0e4710e

note strix-llama.cpp PR #106: qwen4exp decode +19.6% on gfx1151 (MMVQ rows-per-wave, deep prefetch, HIP graph keys, IQ3_S vec-dot); Qwen3.8-Flash-Next

halo-box/strix-llama.cpp#b7149217a12a44bc42974b10497c069d7664df35

note strix-llama.cpp PR #171: server saves the MTP draft context with the slot (prompt-cache restore keeps spec decode warm); djinn/voice prefix caching

halo-box/strix-llama.cpp#c99aae6c2c62a49a98a6b951597a4c57faa8bdf7

note strix-llama.cpp PR #149: server prompt-cache disk spill (owner-only meta); matches strixite card PF-PC disk prompt cache

halo-box/strix-llama.cpp#d7116c8fb818675c4600395ac99604b2280c5ab2

note strix-llama.cpp PR #166: hip/gfx1151 mmap load hang fix, QSA get_rows %32, long-context session stability (measured on Qwen3.8-Flash-Next)

halo-box/strix-llama.cpp#deepseek-v41-core-runtime

note strix-llama.cpp branch adding LLM_ARCH_DEEPSEEK41 (DeepSeek V4.1): loader+graph, engram/expert runtime, expert-store and bounded-file/mmap ple-disk streaming

ikawrakow/ik_llama.cpp

JustVugg/colibri

note NVMe-streaming MoE engine: 744B on 25 GB RAM, slow

kyuz0/strix-halo-ds4-toolbox

leejet/stable-diffusion.cpp

lemonade-sdk/lemonade

note Lemonade (AMD-backed, Apache-2.0, C++): local LLM server + app launcher for Ryzen AI / Radeon, OpenAI API + MCP, Vulkan/ROCm llama.cpp and ONNX-RT/NPU backends; dad joined their Discord 2026-10-13 and wants to contribute (Strix Halo Linux, gfx1151, NPU)

lemonade-sdk/lemonade-mobile

note community mobile client for a Lemonade server

lemonade-sdk/lemonade-vscode

note Lemonade server as the model behind VS Code Copilot Chat

lemonade-sdk/llamacpp-rocm

note lemonade-sdk fresh llama.cpp builds with ROCm 7 per GPU family (gfx1151 among them); release zips pinned in reference.tsv

LostRuins/koboldcpp

mlc-ai/mlc-llm

mostlygeek/llama-swap

MuchDevSuchCode/halo

note Strix Halo XDNA2 NPU: bring-up fix (out-of-tree amdxdna 2.23 built with gcc-15 -> fw 255.0.11.71, 51 TOPS int8), GPU<->NPU zero-copy (hipHostRegister + XRT userptr), open IRON int8 transformer kernels, ggml NPU backend + Gemma-4 MTP segfault patch for atomic-llama; NPU prefill measured 0.31 vs iGPU 329 tok/s

MuchDevSuchCode/NeuralDeck-Proxy

note local LLM control room: OpenAI+Anthropic multimodal proxy routing by model name over llama-server/vllm, dashboard with AMD/NVIDIA telemetry, --help flag probing for forks, per-request benchmarks; NO AUTH and binds 0.0.0.0 by default, bind 127.0.0.1 and tunnel

Nathanw1014/strix-halo-llamacpp

note Nathanw strix-halo-llamacpp: Vulkan llama.cpp tuned for gfx1151 (FA/MoE-prefill/GDN/MTP work), portable tarball bundles Mesa 26.2.4 RADV; dev moving to halo-box/strix-llama.cpp

ollama/ollama

oobabooga/text-generation-webui

open-webui/open-webui

peonist-ai/halogen-flash-server

note halogen docs, docker-compose, deploy/ host units, AGENTS.md and benchmark tools (engine itself is the ghcr image in containers.tsv)

pwilkin/trellis.cpp

note trellis.cpp: TRELLIS.2-4B image-to-3D (BiRefNet->DINOv3->3 flow DiTs->VAEs->UV-textured GLB) in C++/GGML, no Python/CUDA spconv; vendors pwilkin/ggml (branch trellis) as a submodule; Vulkan is fastest on Strix Halo gfx1151; weights = ilintar/trellis2-gguf; libwebp + optional retopo deps via FetchContent (reference.tsv toolchains/trellis.cpp-fetchcontent)

ROCm/FastFlowLM

note FastFlowLM: turnkey LLM runtime for the XDNA2 NPU (kernels closed per halo; ex FastFlowLM/FastFlowLM)

ROCm/hrx-system

note HRX: Hip Runtime Extended (AMD) - source for building the native bundle on older glibc

s-b-repo/peregrine

note Rust spin-off of colibri

shawnshekari/strixite

note strixite: from-scratch HIP engine for Strix Halo (gfx1151) + Qwen3.8-Flash-Next only: ~48 t/s decode flat to 492k context, 512k YaRN, MTP, RAM+disk prompt cache, OpenAI server (no auth, binds 0.0.0.0); builds on TheRock ROCm nightly gfx1151 10.1; weights in its own strixw format (converters from the raw Qwen release in tools/)

SillyTavern/SillyTavern

warpfront/hipfire

note hipfire: Rust+HIP LLM inference for AMD GPUs, tuned first for Strix Halo (gfx1151) with CDNA/RDNA1-4; Qwen3.8-Flash-Next and 27B, DeepSeek V4 Flash, Gemma 4, LFM2; native MTP + DFlash speculation, XDNA2 NPU trunk prefill route, Redline retained-PM4 replay, EP2 on RDNA4; OpenAI-compatible serve; ROCm loaded dynamically (no Python, no PyTorch); Apache-2.0/MIT; weekly Tuesday releases (v0.4.2.1 2026-10-14); reference engine for strixite perf (pod-reports 2026-10-16-hipfire-042-delta-for-strixite.md)

zacharydenton/hrx-rs

note Rust GPU/NPU compute over AMD HRX + Loom (h3-hrx, krea2-hrx family); provisions the native bundle

zacharydenton/hrx-rs#8f0e084a9efd69ec0045b5cd7d032852d048195d

note h3-hrx pin (newer than the 0.8.16 release)

Decision models (Jev-style)17 repos

angel291592/Intent-Router

Mushroom-Systems/lichen

nokia-applied-research/AnyJev

TheoLeeCJ/SemIf-OpenJev

Training, fine-tuning and RL23 repos

Diffusion and video LoRA trainers6 repos

bghira/SimpleTuner

kohya-ss/musubi-tuner

kohya-ss/sd-scripts

Nerogar/OneTrainer

ostris/ai-toolkit

tdrussell/diffusion-pipe

Quantisation and merging6 repos

arcee-ai/mergekit

huggingface/optimum

huggingface/optimum-quanto

ModelCloud/GPTQModel

pytorch/ao

turboderp-org/exllamav3

Conversion and abliteration3 repos

efzero/diffusioncontroller

p-e-w/heretic

vllm-project/llm-compressor

Creative apps (ArtCraft Craft suite)29 repos

storytold/AniCraft

note AniCraft: 2D animation editor in pure Rust (ArtCraft; new 2026-10, no licence file yet)

storytold/artcraft

note ArtCraft: the ArtCraft org's (storytold) 'intentional crafting engine' for artists, designers and filmmakers; pure Rust; Apache-2.0; 1.3 GB repo (assets in tree); hub of the Craft apps below; dad 2026-10-16

storytold/artcraft-services

note artcraft-services: ArtCraft backend and web frontend (Rust; 1.3 GB; no licence file, NOASSERTION: source access only, check before reuse)

storytold/artcraftx

note ArtCraft-X (Rust, 37 MB, no licence file)

storytold/cadcraft

note CADCraft: clean-room AutoCAD-style 2D CAD and drafting in pure Rust (ArtCraft, Apache-2.0)

storytold/craft-common-actions

note craft-common-actions: shared GitHub Actions (Python) the Craft apps build with; source of their Linux build-dependency list

storytold/craft-contrib

note craft-contrib: ArtCraft contributor material (new 2026-10-15)

storytold/craft-fonts

note craft-fonts: fonts bundled by the Craft apps (36 MB, Apache-2.0)

storytold/craft-launcher

note craft-launcher: the Craft App Launcher (installs/updates the apps) in Rust (ArtCraft, Apache-2.0)

storytold/craft-launcher-assets

note craft-launcher-assets: screenshots the launcher downloads at run time (details pages); mirrored so the launcher works offline

storytold/craft-libs

note craft-libs: shared Rust libraries for the Craft apps (Apache-2.0); the apps pin it as a git dependency by rev, so the matching revs live in cargo-home/git

storytold/craft-repo-status-app

note craft-repo-status-app: PR/build status board for the Craft repos (Rust)

storytold/crafthub

note crafthub: ArtCraft hub repo (new 2026-10-16, near-empty; tracked as it fills)

storytold/deckcraft

note DeckCraft: presentations and slide shows, clean-room PowerPoint reimplementation in pure Rust (ArtCraft, Apache-2.0)

storytold/designcraft

note DesignCraft: page and layout design editor (InDesign-style) in pure Rust (ArtCraft, Apache-2.0)

storytold/effectcraft

note EffectCraft: motion graphics and compositing (After Effects-style) in pure Rust (ArtCraft, Apache-2.0)

storytold/figcraft

note FigCraft: interface design and prototyping, clean-room Figma-style collaborative design in pure Rust (ArtCraft, Apache-2.0)

storytold/filmcraft

note FilmCraft: nonlinear video editor in pure Rust (ArtCraft, Apache-2.0); git dep craft-libs

storytold/gridcraft

note GridCraft: clean-room Excel-style spreadsheet in pure Rust (ArtCraft, Apache-2.0)

storytold/indd-utils

note indd-utils: read Adobe INDD files without InDesign and convert to IDML (Rust, Apache-2.0); DesignCraft import path

storytold/lightcraft

note LightCraft: non-destructive photo editor (Lightroom-style) in pure Rust (ArtCraft, Apache-2.0)

storytold/pdfcraft

note PdfCraft: PDF reader and editor in pure Rust (ArtCraft, Apache-2.0)

storytold/photocraft

note PhotoCraft: open-source photo editor from scratch in pure Rust (ArtCraft Craft app, Apache-2.0); needs craft-libs and the storytold winprint.rs fork at pinned revs (git deps, in cargo-home)

storytold/photocraft-corpus

note photocraft-corpus: test images and data for PhotoCraft (Apache-2.0)

storytold/soundcraft

note SoundCraft: clean-room Avid Pro Tools reimplementation (DAW) in pure Rust (ArtCraft, Apache-2.0)

storytold/vectorcraft

note VectorCraft: vector art tool (Illustrator-style) in pure Rust (ArtCraft, Apache-2.0)

storytold/winit

note storytold fork of winit (Apache-2.0, master): the window layer the Craft apps patch; kept for their fork-specific fixes

storytold/winprint.rs

note storytold fork of winprint.rs (BSD-3-Clause): pinned git dependency of craft-libs and photocraft

storytold/wordcraft

note WordCraft: clean-room Microsoft Word reimplementation in pure Rust (ArtCraft, Apache-2.0)

Evaluation and data tooling21 repos

AAAndy-Zhu/VGCure

note VGCure (ACL 2025): 22-task visual graph understanding benchmark; edge counting and neighbour tasks are where VLMs collapse (best open model 16% on edge count), plus MCDGraph self-supervised fix; baseline for visual-graph memory edge attribution

argilla-io/distilabel

camilochs/visgraphvar

note VisGraphVar: generator for visual-graph benchmarks with controlled variability (layout, style, size); synthetic node-link graphs with ground truth for edge-attribution sweeps

EleutherAI/lm-evaluation-harness

HITsz-TMG/VisionGraph

note VisionGraph: 17.6k questions over 5.8k rendered graphs, 8 graph-theory tasks + node/edge perception probes; dataset MrSunshy/VisionGraph_GraphStructure_hf

huggingface/datasets

huggingface/datatrove

huggingface/evaluate

huggingface/lighteval

huggingface/smol-course

Jo-Pan/FlowLearn

note FlowLearn: 3.9k scientific + 10k simulated flowcharts with Mermaid code, OCR and VQA annotations; edge-level F1 for frontier VLMs is ~0.2-0.3, the number to beat; dataset jopan/FlowLearn

khimaros/aimbot

note model-facts registry + ranked comparison for a 128 GB Strix Halo (GPL-3.0): registry/models.yaml (quants, context, thinking knobs, sampling profiles), docs/data.json + scripts/aimbot roster CLI run offline

khimaros/quantbench

note run terminal-bench against llama.cpp quants (the measurements behind aimbot)

MiaSanLei/MindBench

note MindBench (Meituan, 2024): mind-map structure recognition benchmark, ~23k crawled XMind/Biggerplate maps + ~400k Graphviz-synthesised maps over 5 tasks (full/part/position parsing, structured VQA); parsing + synthesis + eval code; CC BY-NC 4.0; dataset MiaSanLei/MindBench

Moenupa/VTCBench

note VTCBench: can VLMs understand long context under vision-text compression (retrieval, reasoning, memory settings); the public benchmark closest to OMF H2; dataset MLLM-CL/VTCBench

NishilBalar/Awesome-LVLM-Hallucination

note Awesome LVLM Hallucination: curated list incl. OCR hallucination under degradation (KIE-HVQA, abstention probes); offline index for the smudge abstain-vs-confabulate question

nju-websoft/FlowGen

note FlowGen (ICLR 2026): controllable synthetic flowchart generator with Mermaid/Graphviz/PlantUML/Diagrams backends; raises Qwen2.5-VL-7B 74.6->89.2 on FlowVQA; dataset Anonymous112233/FlowGen

opendatalab/OmniDocBench

note OmniDocBench v1.5: the document-parsing benchmark every OCR model above reports on

UKGovernmentBEIS/inspect_ai

WEIYanbin1999/GITA

note GITA (NeurIPS 2024): graph-to-visual+textual integration, GVLQA dataset, rendering-level augmentation (layout, node shape, edge thickness); shows visual+text beats either alone on graph reasoning

zhaochen0110/Awesome_Think_With_Images

note Awesome Think With Images: survey + paper/code list for VLM zoom/crop/search tool use (DeepEyes, Chain-of-Focus, V*, ZoomEye...); offline index for the zoom-tool literature

Multi-box serving and clustering15 repos

b4rtaz/distributed-llama

BerriAI/litellm

cmetz/ec-su_axb35-linux

note ec_su_axb35: out-of-tree Linux driver for the Sixunited AXB35-02 EC (Bosgame M5, GMKtec EVO-X2, ...): sysfs power_mode quiet/balanced/performance (EC 0x31), fan1-3 rpm/mode/level/curves, EC temp; the Pod workers have no other fan or power-mode lever (rebuild per kernel; plain make against the headers)

cmetz/ec-su_axb35-linux#f62c2c228959a08683273a26ef3afd8991e69f6d

note PINNED 2026-10-05 commit: the one built into the Pod worker image (7.0.0-34)

exo-explore/exo

Geramy/OdinLink-Five

note OdinLink: Thunderbolt 5/USB4 as an RDMA interconnect (kernel module + ibv_* verbs provider + RCCL/NCCL net plugin); vLLM TP=2 on two Strix Halos +20-26% tok/s over TCP-over-TB; Strix Halo 20/40G negotiation notes in issue #31

hellas-ai/thunderbolt-ibverbs

note upstream of donnerkeule: RDMA (ibverbs) over USB4/Thunderbolt, without write striping

neuhaus/donnerkeule#feat/write-striping

note thunderbolt_ibverbs kernel module + userspace: USB4/Thunderbolt DMA rings as RoCE v2 RDMA devices (usb4_rdma*); write-striping branch spreads one QP over every ring/cable (2x40G cables ~44 Gbit/s); what gufo TP2 uses

neuhaus/ds4-strix-halo-tp-odinlink

note neuhaus fork of the above (also InfiniBand); same author as the gufo RDMA fork

neuhaus/gufo

note Gufo RDMA fork (default branch rdma): two-Strix-Halo tensor parallel (TP2) over InfiniBand, RoCE v2 or USB4 cables; Qwen3.8 Flash-Next Q4/Q8; build -DGUFO_ENABLE_TP2_RDMA=ON (libibverbs); docs/models/qwen3.8-flash-next/TP2.md

OpenCPIL/prima.cpp

sgl-project/sglang

vllm-project/vllm

wkljohn/ds4-strix-halo-tp-odinlink

note DeepSeek V4 Flash tensor parallel on two Strix Halos over OdinLink USB4/TB5 RDMA or Mellanox RoCE v2

wkljohn/ds4-strix-halo-tp-odinlink#71e6a2475952c0dd49677a8dcc42661854047199

note PINNED 2026-09-15 commit: last one that builds on ROCm 7.x (HEAD needs ROCm 10); the Pod's DeepSeek V4 Flash TP=2 build (ROCm 7.2.4, DS4_ALLOW_ROCM_MISMATCH=1, RDMA_PROFILE=roce-v2 over usb4_rdma)

ML frameworks and libraries39 repos

abcfy2/aria2-static-build

aria2/aria2

astral-sh/uv

containers/podman

docker/docker-install

huggingface/diffusers

huggingface/huggingface_hub

huggingface/safetensors

huggingface/tokenizers

huggingface/transformers

k2-fsa/sherpa-onnx

KhronosGroup/glslang

KhronosGroup/SPIRV-Headers

KhronosGroup/SPIRV-Tools

KhronosGroup/Vulkan-Headers

KhronosGroup/Vulkan-Loader

KhronosGroup/Vulkan-Tools

microsoft/onnxruntime

ROCm/composable_kernel

ROCm/rocm-install-on-linux

ROCm/TheRock

note TheRock: AMD's open ROCm build system and the source of the nightly/stable ROCm tarballs (repo.amd.com/rocm/core/tarball); strixite builds against its gfx1151 nightly (see reference.tsv toolchains/therock-gfx1151)

Model source repos33 repos

ace-step/ACE-Step

allenai/olmocr

note olmOCR pipeline + olmOCR-bench

baidu/unlimited-ocr

note Unlimited-OCR (Baidu, 2026): DeepSeek-OCR lineage, one-shot long-horizon document parsing (dozens of pages per forward pass, R-SWA constant KV); the optical-context-compression encoder/decoder the OMF experiment ports to ROCm; CUDA-only release stack

black-forest-labs/flux2

deepseek-ai/DeepSeek-OCR

note DeepSeek-OCR v1 code + vLLM recipe

deepseek-ai/DeepSeek-OCR-2

facebookresearch/dinov2

note DINOv2 reference code (imageboop inject/refbank.py parity; CI-5)

facebookresearch/sam-3d-body

facebookresearch/sam-3d-objects

FunAudioLLM/CosyVoice

HeartMuLa/heartlib

index-tts/index-tts

jasongzy/Make-It-Animatable

note MinerU document parsing pipeline

PaddlePaddle/PaddleOCR

QwenLM/Qwen-Image

QwenLM/Qwen3-ASR

QwenLM/Qwen3-TTS

rednote-hilab/dots.ocr

note dots.ocr code

resemble-ai/chatterbox

Tencent-Hunyuan/Hunyuan3D-2

Tencent-Hunyuan/Hunyuan3D-2.1

TencentARC/Pixal3D

thu-coai/Glyph

note Glyph (THU CoAI + Zhipu): renders long text to images and reads it with a VLM, 3-4x token compression on GLM-4.1V; the general-VLM alternative to a dedicated OCR decoder for optical memory

Ucas-HaoranWei/GOT-OCR2.0

note GOT-OCR2.0 code

VAST-AI-Research/UniRig

Wan-Video/Wan2.2

yuripourre/Pixal3D-ROCm

zai-org/GLM-OCR

note GLM-OCR code + tech report

ComfyUI and custom nodes62 repos

1038lab/ComfyUI-QwenTTS

1038lab/ComfyUI-RMBG

Acly/comfyui-inpaint-nodes

Acly/comfyui-tooling-nodes

AIGODLIKE/AIGODLIKE-ComfyUI-Translation

akatz-ai/ComfyUI-Depthflow-Nodes

BadCafeCode/masquerade-nodes-comfyui

benjiyaya/HeartMuLa_ComfyUI

chflame163/ComfyUI_LayerStyle

chrisgoringe/cg-use-everywhere

christian-byrne/audio-separation-nodes-comfyui

city96/ComfyUI-GGUF

city96/ComfyUI_ExtraModels

ClownsharkBatwing/RES4LYF

Comfy-Org/comfy-cli

Comfy-Org/ComfyUI-Manager

Comfy-Org/ComfyUI_frontend

comfyanonymous/ComfyUI

crystian/ComfyUI-Crystools

cubiq/ComfyUI_essentials

diodiogod/TTS-Audio-Suite

Fannovel16/ComfyUI-MotionDiff

filliptm/ComfyUI_Fill-Nodes

Gourieff/ComfyUI-ReActor

if-ai/ComfyUI-IF_Trellis

jags111/efficiency-nodes-comfyui

Jonseed/ComfyUI-Detail-Daemon

kijai/ComfyUI-DepthAnythingV2

kijai/ComfyUI-FramePackWrapper

kijai/ComfyUI-GIMM-VFI

kijai/ComfyUI-Hunyuan3DWrapper

kijai/ComfyUI-HunyuanVideoWrapper

kijai/ComfyUI-KJNodes

kijai/ComfyUI-LivePortraitKJ

kijai/ComfyUI-LTXVideo

kijai/ComfyUI-MMAudio

kijai/ComfyUI-segment-anything-2

kijai/ComfyUI-SUPIR

kijai/ComfyUI-WanAnimatePreprocess

kijai/ComfyUI-WanVideoWrapper

Kosinkadink/ComfyUI-Advanced-ControlNet

Lightricks/ComfyUI-LTXVideo

lldacing/ComfyUI_BiRefNet_ll

logtd/ComfyUI-Fluxtapoz

ltdrdata/ComfyUI-Impact-Pack

ltdrdata/ComfyUI-Inspire-Pack

MrForExample/ComfyUI-3D-Pack

nullquant/ComfyUI-BrushNet

patientx/ComfyUI-INT8-Fast-ROCM

PowerHouseMan/ComfyUI-AdvancedLivePortrait

PozzettiAndrea/ComfyUI-SAM3DObjects

pythongosssss/ComfyUI-Custom-Scripts

rgthree/rgthree-comfy

ShmuelRonen/ComfyUI-VideoUpscale_WithModel

spacepxl/ComfyUI-Image-Filters

ssitu/ComfyUI_UltimateSDUpscale

stavsap/comfyui-ollama

Suzie1/ComfyUI_Comfyroll_CustomNodes

visualbruno/ComfyUI-Trellis2

welltop-cn/ComfyUI-TeaCache

yolain/ComfyUI-Easy-Use

yuvraj108c/ComfyUI-Whisper

Image, video and vision31 repos

cubiq/ComfyUI_IPAdapter_plus

danielgatis/rembg

deepinsight/insightface

DepthAnything/Depth-Anything-V2

facebookresearch/sam2

Fanghua-Yu/SUPIR

Fannovel16/ComfyUI-Frame-Interpolation

Fannovel16/comfyui_controlnet_aux

HiDream-ai/HiDream-I1

hkchengrex/MMAudio

kijai/ComfyUI-Florence2

Kosinkadink/ComfyUI-AnimateDiff-Evolved

Kosinkadink/ComfyUI-VideoHelperSuite

lllyasviel/FramePack

ltdrdata/ComfyUI-Impact-Subpack

mu-cai/matryoshka-mm

note Matryoshka Multimodal Models (ICLR 2025): nested visual token sets, coarse derived from fine, token count chosen at inference; the multi-resolution fidelity knob idea for a resolution ladder

multimodal-art-projection/YuE

om-ai-lab/ZoomEye

note ZoomEye: training-free tree-search zoom for any VLM (each child node is a sub-region); the progressive-resolution reader loop for optical memory tiers, slow (~50 s/question) but model-agnostic

PrismML-Eng/Bonsai-Image-Demo

note local image generation demo (upstream of the MuchDevSuchCode fork)

pyannote/pyannote-audio

sczhou/CodeFormer

suno-ai/bark

Tencent-Hunyuan/HunyuanVideo-1.5

TencentARC/GFPGAN

Tongyi-MAI/Z-Image

ultralytics/ultralytics

Visual-Agent/DeepEyes

note DeepEyes (ICLR 2026): RL-trained zoom/crop tool use inside the reasoning trace (verl, PPO/GRPO); the trained alternative to ZoomEye; weights ChenShawn/DeepEyes-7B

WASasquatch/was-node-suite-comfyui

xinntao/Real-ESRGAN

zacharydenton/h3-hrx

note MiniMax H3 video+audio on Strix Halo gfx1151 in Rust (Loom kernels + HRX, no Python/PyTorch/ROCm SDK); needs glibc 2.43 bundle (shim/26.04) + int8_convrot DiTs

ZhengPeng7/BiRefNet

Media tooling (audio, video, image)17 repos

3b1b/manim

audacity/audacity

facebookresearch/audiocraft

facebookresearch/demucs

FFmpeg/FFmpeg

GNOME/gimp

ImageMagick/ImageMagick

MuchDevSuchCode/Galileo

note WinUI 3 / .NET 8 file manager + photo viewer (Windows Explorer/Photos alternative)

remotion-dev/remotion

spotify/pedalboard

Stability-AI/stable-audio-tools

Audio editing and restoration13 repos

declare-lab/tango

haoheliu/AudioLDM2

haoheliu/voicefixer

jthickstun/anticipation

m-bain/whisperX

MontrealCorpusTools/Montreal-Forced-Aligner

Plachtaa/seed-vc

resemble-ai/resemble-enhance

Rikorose/DeepFilterNet

RVC-Project/Retrieval-based-Voice-Conversion-WebUI

sergree/matchering

surge-synthesizer/surge

ZFTurbo/Music-Source-Separation-Training

Voice assistant stack37 repos

canopyai/Orpheus-TTS

dscripka/openWakeWord

fishaudio/fish-speech

fixie-ai/ultravox

getlark/openlily

hexgrad/kokoro

huggingface/speech-to-speech

idiap/coqui-ai-TTS

KoljaB/RealtimeSTT

KoljaB/RealtimeTTS

kyutai-labs/delayed-streams-modeling

kyutai-labs/moshi

kyutai-labs/pocket-tts

kyutai-labs/unmute

livekit/agents

livekit/livekit-wakeword

mistralai/mistral-common

moonshine-ai/moonshine

pipecat-ai/smart-turn

QwenLM/Qwen3-Omni

rhasspy/wyoming

rhasspy/wyoming-faster-whisper

rhasspy/wyoming-openwakeword

rhasspy/wyoming-piper

rhasspy/wyoming-satellite

SesameAILabs/csm

shawnshekari/navi-tts

note navi-tts: Qwen3-TTS 0.6B in hand-written HIP for gfx1100/gfx1151 (same author as strixite); OpenAI + XTTS APIs, voice cloning; Strix RTF 0.27, ~2.7 GiB; builds with TheRock ROCm >=10.1 clang; AGPL-3.0; no auth (LAN only)

snakers4/silero-vad

speaches-ai/speaches

stepfun-ai/Step-Audio2

SYSTRAN/faster-whisper

thewh1teagle/kokoro-onnx

zai-org/GLM-4-Voice

Agent harnesses and frameworks47 repos

agent0ai/agent-zero

agentpatternscatalog/patterns

note Agent Patterns Catalog (agentpatternscatalog.org): ~530 agent/LLM design patterns + framework compositions, markdown + JSON, CC BY 4.0

Aider-AI/aider

All-Hands-AI/OpenHands

anthropics/claude-agent-sdk-python

anthropics/claude-code

block/goose

browser-use/browser-use

cline/cline

continuedev/continue

crewAIInc/crewAI

deepseek-ai/deepseek-harness

earendil-works/pi

google-gemini/gemini-cli

google/adk-python

gptme/gptme

huggingface/smolagents

Kilo-Org/kilocode

langchain-ai/langgraph

langflow-ai/langflow

langgenius/dify

lemonade-sdk/peel

note local-LLM aid inside PowerShell (Lemonade client)

modelcontextprotocol/python-sdk

modelcontextprotocol/servers

modelcontextprotocol/typescript-sdk

MuchDevSuchCode/Prompt-Factory

note desktop app: plain-language idea -> engineered prompt for 50 AI tools, local LLM or cloud API

n8n-io/n8n

NousResearch/hermes-agent

openai/codex

openai/openai-agents-python

openclaw/openclaw

openinterpretability/ekbasis

note Ekbasis: calibrated consequence model for agents (git/shell guard before execution, Claude Code PreToolUse hook, MCP, look-when-unsure simulator); serves on vLLM>=0.30 letter-logprob readout; 2 Zenodo papers, pre-registered evals

openinterpreter/open-interpreter

overmind-core/overmind

note Overmind: agent training platform: repo capability graph, OTel trace scoring, trace->dataset, evals, prompt/tool/control-flow optimisers (winner = git diff), fine-tune + backtest on own traces, MCP server; platform AGPL-3.0, SDK MIT; self-host needs OpenRouter/Modal/AWS keys (not offline as-is)

Picrew/awesome-agent-harness

pydantic/pydantic-ai

QwenLM/qwen-code

RooCodeInc/Roo-Code

run-llama/llama_index

Significant-Gravitas/AutoGPT

sst/opencode

SWE-agent/mini-swe-agent

SWE-agent/SWE-agent

Multi-agent and swarm frameworks28 repos

Agent orchestration and workflows16 repos

2FastLabs/agent-squad

AMD-AGI/GEAK

note GEAK: AMD kernel-generation agent backend that Hyperloom delegates hot kernels to (Triton/HIP)

AMD-AGI/Hyperloom

note ROCm Hyperloom (MIT, v1.1.3): multi-agent harness that auto-optimizes LLM inference on AMD GPUs (prelude baseline+roofline -> framework flags/diffs -> kernel agent -> sweep; recipe knowledge base; bounded edits). dad 2026-10-12; sibling of the overmind epic and strixite/dsv4 ROCm tuning

langchain-ai/deepagents

langchain-ai/langchain

open-multi-agent/open-multi-agent

stanfordnlp/dspy

Untrivial-ai/agent-orchestrator

Agent protocols (MCP, A2A, AG-UI)11 repos

a2aproject/A2A

a2aproject/a2a-python

ag-ui-protocol/ag-ui

agent-network-protocol/AgentNetworkProtocol

agentpatternscatalog/mcp

note MCP server for the Agent Patterns Catalog; offline: --catalog-dir <mirror>/src/github.com__agentpatternscatalog__patterns (else it fetches from GitHub on first run); no licence file

anthropics/anthropic-sdk-python

anthropics/claude-agent-sdk-typescript

jlowin/fastmcp

lastmile-ai/mcp-agent

mcp-use/mcp-use

punkpeye/fastmcp

Parallel coding agents and skills13 repos

anthropics/skills

BloopAI/vibe-kanban

humanlayer/humanlayer

microsoft/playwright-mcp

MuchDevSuchCode/OpenGravity

note VS Code coding assistant on local LLMs (Antigravity alternative)

stagewise-io/stagewise

SWE-agent/SWE-ReX

Self-hosted fronts, guardrails and eval8 repos

caramaschiHG/awesome-ai-agents-2026

danny-avila/LibreChat

EvoMap/awesome-agent-swarm

lobehub/lobe-chat

Mintplex-Labs/anything-llm

NVIDIA-NeMo/Guardrails

princeton-nlp/SWE-bench

THUDM/AgentBench

Retrieval and documents16 repos

chroma-core/chroma

datalab-to/marker

docling-project/docling

facebookresearch/faiss

FlagOpen/FlagEmbedding

UKPLab/sentence-transformers

Unstructured-IO/unstructured

Island dev infrastructure13 repos

emacs-mirror/emacs

go-gitea/gitea

graphviz/graphviz

note Graphviz: dot/neato/sfdp layouts; the reference renderer for node-link graphs and a FlowGen backend

jgm/pandoc

jupyterlab/jupyterlab

marimo-team/marimo

mermaid-js/mermaid-cli

note mermaid-cli: headless Mermaid renderer (node/puppeteer); renders synthetic graphs and flowcharts for VLM memory experiments and FlowGen/FlowLearn-style data

MuchDevSuchCode/FTS-Network-Monitor

note dual-WAN monitor: ICMP/TCP loss, latency, jitter to router/gateways/DNS with a local web UI

MuchDevSuchCode/KeyWeaver

note two-passphrase deterministic cryptographic key generator

neovim/neovim

squidfunk/mkdocs-material

terrastruct/d2

note d2: declarative diagram language and renderer (Go, no browser) with multiple layout engines; alternative graph renderer for node-link memory pages

typst/typst

Network boot and node provisioning12 repos

canonical/cloud-init

note NoCloud datasource reference for autoinstall user-data/meta-data

canonical/curtin

note the installer's storage/curtin layer that autoinstall drives

canonical/subiquity

note Ubuntu Server installer + autoinstall schema source (ships as a snap inside the live ISO, not in apt)

canonical/ubuntu-image

note Canonical's image builder (image definitions, gadget trees)

danderson/netboot

note pixiecore: one static Go binary = proxyDHCP + TFTP + HTTP PXE server; boots nodes on a segment with no DHCP server

dracut-ng/dracut-ng

note initramfs with network root (nfs/iscsi/nbd/livenet); dracut-network deb is in the apt mirror

ipxe/ipxe

note iPXE: build a custom UEFI netboot binary with an embedded script (snponly.efi for Realtek RTL8125 via firmware SNP); the Ubuntu ipxe deb is in the apt mirror

ltsp/ltsp

note diskless clients booting a squashfs image over NFS/NBD (ltsp deb is in the apt mirror)

netbootxyz/netboot.xyz

note iPXE menu system; usable offline when its assets are mirrored locally

systemd/mkosi

note build bootable OS images (UKI, squashfs, disk) from the apt mirror without a live ISO

ventoy/Ventoy

note multi-ISO USB stick: the no-network fallback for installing a node

warewulf/warewulf

note stateless image-based cluster provisioning (overlay + container images over PXE); the HPC way to run a diskless GPU fleet

Developer toolchains (sources)19 repos

amd/xdna-driver

note amdxdna NPU driver (XDNA/XDNA2); on Strix Halo build 2.23 out-of-tree with gcc-15 to match XRT 2.23 (see halo scripts/npu-bringup.sh)

note Playwright: drives Chromium/WebKit/Firefox headless for page tests (device emulation, CDP touch, network throttling); its browser builds are downloads, not derivable from the repo

note Peano: LLVM backend for AIE cores; wheels on releases/nightly

Xilinx/mlir-aie

note IRON / MLIR-AIE: open AIE kernel toolchain (aie2p = XDNA2); wheels on releases/latest-wheels-2

Xilinx/XRT

note Xilinx Runtime: NPU userspace (xrt-smi, pyxrt, userptr buffers)

Offline reference tooling4 repos

GPUOpen-Tools/isa_spec_manager

note AMD machine-readable ISA XML (CDNA1-4, RDNA1-4, RDNA3.5) + parser; basis of the cdna4-isa/rdna35 shelves

kiwix/kiwix-desktop

kiwix/kiwix-tools

openzim/zim-tools

Embedded firmware12 repos

ablueway/fullmac

note iComm/SSV fullmac WiFi host SDK (SSV6030/SSV6006) including the Generalplus GP15B platform tree (GPL326xx ARM920T, uC/OS-II: driver_l1/l2, gplib, UVC usbd, SSV porting layer) with CodeSourcery/arm-none-eabi make rules; the SDK behind the WiFi microscope custom firmware (dad 2026-10-14, see Code/wifi-microscope/FIRMWARE.md)

arduino/arduino-cli

earlephilhower/arduino-pico

espressif/arduino-esp32

espressif/esp-idf

freerouting/freerouting

Links2004/arduinoWebSockets

note arduinoWebSockets (Arduino lib 'WebSockets'): WebSocket server/client for ESP32/ESP8266 Arduino cores; WEBSOCKETS_SERVER_CLIENT_MAX must be set as a build flag (array size in a header), not in the sketch

lionyhw/JHDH-8412_1247

note JoyHonest GPCV1248 camera board dump: startup.s, release.scf, SPII SDRAM boot header, schematic PDF, G+CodePacker + USB mass-production tool; supplies the startup code fullmac lacks (WiFi microscope firmware)

platformio/platformio-core

qmk/qmk_firmware

raspberrypi/pico-sdk

wythe-lin/ZTKVR

note Generalplus GPL326xx (GP6813) car-DVR SDK: driver_l1/l2, gplib incl. AviPackerV3, project + tool dirs; second reference copy of the GP326 platform for the WiFi microscope firmware

Edge vision (Grove Vision AI V2)15 repos

google-ai-edge/LiteRT

HimaxWiseEyePlus/himax_tflm

HimaxWiseEyePlus/Seeed_Grove_Vision_AI_Module_V2

note Grove Vision AI V2 firmware SDK (Himax WiseEye2)

IntelRealSense/librealsense

note RealSense D4xx depth+RGB driver (librealsense2, rsusb backend, udev rules); djinn DJ-7 camera on pod-control

nxp-imx/ethos-u-vela

note Ethos-U55 compiler

OpenKinect/libfreenect

note Kinect v1 driver (firmware loader pulls it off the sensor); djinn DJ-7 alternative camera

OpenKinect/libfreenect2

note Kinect v2 driver (depth+RGB over USB3); djinn DJ-7 alternative camera

PINTO0309/onnx2tf

note ONNX -> TFLite

Seeed-Studio/ModelAssistant

note SSCMA: train models for the Grove

Seeed-Studio/Seeed_Arduino_SSCMA

note XIAO I2C client

Seeed-Studio/SenseCraft-Web-Toolkit

note browser flasher

Seeed-Studio/sscma-example-we2

Seeed-Studio/sscma-micro

note runtime firmware

Seeed-Studio/sscma-model-zoo

tensorflow/tflite-micro

Games16 repos

KhronosGroup/glTF-Sample-Assets

libsdl-org/SDL

love2d/love

mapeditor/tiled

Mesh2Motion/mesh2motion-app

note Mesh2Motion (MIT): browser auto-rigger + animation library, the open Mixamo replacement; import GLB/GLTF/DAE/FBX, pick a skeleton (human, horse, fox, bird, dragon, spider, snake, fish, kaiju), fit it, export bundled animations as GLB/FBX; 70 MB of CC0 animations/rigs/previews are in-repo under static/; TS + Vite 8, Node 24 (.nvmrc); run `npm ci` in the clone on x86-64 so node_modules ships (the lockfile carries wrangler/workerd/sharp for Cloudflare deploys, ~100 MB, harmless). Off-island surprises: styles.css pulls the Jost font from fonts.googleapis.com (falls back to sans-serif; the OFL file is in google/fonts ofl/jost), and the download/animation-request dialogs POST telemetry to mesh2motion-app.scottpetrovic.workers.dev (fails silently). Full kit: docs/MESH2MOTION.md

Mesh2Motion/mesh2motion-assets

note Mesh2Motion Blender sources (CC0, ~385 MB): rigs, animations and sample models behind mesh2motion-app; needed to add or retarget animations on the island

Mesh2Motion/mesh2motion-desktop

note Mesh2Motion desktop wrapper: the offline build of the app for machines without a dev toolchain

phaserjs/phaser

pygame/pygame

raysan5/raylib

3D modelling and capture23 repos

aframevr/aframe

note A-Frame: declarative WebXR/WebGL scenes on three.js (super-three fork); dist/aframe-v*.min.js is self-contained (three bundled, ~330 KB gz) so a scene runs from a microcontroller with no CDN; rickroll-ap's council chambers

alicevision/Meshroom

assimp/assimp

CadQuery/cadquery

cnr-isti-vclab/meshlab

cnr-isti-vclab/PyMeshLab

colmap/colmap

donmccurdy/glTF-Transform

note glTF-Transform: CLI + library to optimise glTF (meshopt/draco geometry, texture resize/webp/ktx2, dedup, prune); shrinks baked Blender exports for flash-sized web scenes

elalish/manifold

FreeCAD/FreeCAD

graphdeco-inria/gaussian-splatting

gumyr/build123d

isl-org/Open3D

mikedh/trimesh

nerfstudio-project/nerfstudio

nmwsharp/polyscope

Open-Cascade-SAS/OCCT

openscad/openscad

pmp-library/pmp-library

projects.blender.org/blender/blender

prusa3d/PrusaSlicer

SoftFever/OrcaSlicer

zeux/meshoptimizer

note meshoptimizer: mesh simplification/compression; EXT_meshopt_compression decoder (~20 KB js) that three.js/A-Frame load glTF through; gltfpack CLI

Circuit design and HDL25 repos

Genome analysis27 repos

BCI, EEG and biosignals26 repos

reverse-engineering3 repos

czietz/wifimicroscope

note JHCMD UDP protocol proof of concept for the JoyHonest/Generalplus WiFi microscope (first public write-up of the protocol)

gregsadetsky/wifi-endoscope-hd-1200p

note JHCMD protocol + vendor APK findings for the sibling JoyHonest WiFi Check endoscope (GP4225); dump + OpenCV viewer scripts

morluto/rea

note agentic reverse engineering (app behaviour down to native binaries; TypeScript, MIT); pairs with the device-workshop

Model weights190

Picks follow a newest-only policy: one current model per line (the Sept 2026 set: DeepSeek V4.1, GLM-5.3, Kimi K2.7, Qwen3.8, MiMo V2.6, Nemotron 3 Ultra, Ornith 1.5, Mistral Medium 3.5, Muse-Glimmer, plus decision models). Dependencies are not picks: they are the exact checkpoints that mirrored code loads at first run (tokenizer base models, CLIP/SigLIP encoders, whisper, bge, Comfy node weights), so some are deliberately old. include = the glob patterns actually fetched.

Model picks: raw releases (newest of each line; re-quant source)66

allenai/olmOCR-2-7B-1025

note olmOCR-2 7B (AI2, Oct 2025): PDF-to-text VLM, RL-tuned, olmOCR-bench reference (arXiv 2502.18443)

baidu/Unlimited-OCR

note Unlimited-OCR 3B (Baidu, 2026): optical context compression encoder+decoder, one-shot multi-page parsing; the model the OMF experiment ports to ROCm (MIT-style, check LICENSE)

caiovicentino1/Ekbasis-27B

note Ekbasis-27B bf16 (55 GB; Qwen3_5 hybrid GDN, from Qwen3.8-27B via Eikos-27B): source for a GGUF port (llama.cpp letter-logprob readout, recalibrate vs RELEASE_EVAL)

caiovicentino1/Ekbasis-27B-INT4

note Ekbasis-27B INT4 (18.6 GB, compressed-tensors pack-quantized g128; vLLM build for 32 GB GPUs)

ChenShawn/DeepEyes-7B

note DeepEyes 7B: Qwen2.5-VL-7B RL-trained to crop/zoom during reasoning; zoom-tool reader candidate

convaiinnovations/laya

note decision model

crh225/plumb-4b

note decision model

deepseek-ai/DeepSeek-OCR

note DeepSeek-OCR (Oct 2025): contexts optical compression, DeepEncoder + 3B MoE decoder; the paper behind the OMF premise (arXiv 2510.18234)

deepseek-ai/DeepSeek-V4.1-Flash

include .gitattributes DeepSeek_V41_Tech_Report.pdf LICENSE encoding/encoding.py encoding/test_encoding.py evaluation/dsh-minimal.patch inference/convert.py inference/engram.py inference/generate.py inference

facebook/nougat-base

note Nougat (arXiv 2308.13418): academic PDF to markup, Swin encoder + mBART decoder - OCR-free baseline

FastFlowLM/Embedding-Gemma-300M-NPU2

include *.q4nx *.xclbin *.json *.md note FastFlowLM EmbeddingGemma-300M for the XDNA2 NPU (flm embed-gemma:300m, 0.65 GB): NPU-native embeddings

FastFlowLM/Llama-3.2-1B-NPU2

include *.q4nx *.xclbin *.json *.md note FastFlowLM q4nx weights + xclbins for the XDNA2 NPU (flm llama3.2:1b, 1.3 GB): NP-2 job 4

FastFlowLM/Qwen3-1.7B-NPU2

include *.q4nx *.xclbin *.json *.md note FastFlowLM q4nx weights + xclbins for the XDNA2 NPU (flm qwen3:1.7b, 1.7 GB): NP-2 job 4, <=4B LLM on the Strix Halo NPU

FastFlowLM/Qwen3-4B-Instruct-2507-NPU2

include *.q4nx *.xclbin *.json *.md note FastFlowLM q4nx weights + xclbins for the XDNA2 NPU (flm qwen3-it:4b, 3.3 GB): NP-2 job 4

FastFlowLM/Whisper-V3-Turbo-NPU2

include *.q4nx *.xclbin *.json *.md note FastFlowLM Whisper large-v3-turbo for the XDNA2 NPU (flm whisper-v3:turbo, 0.65 GB): NP-2 job 2, STT on the NPU

ggerganov/whisper.cpp

include ggml-large-v3-turbo*.bin ggml-base.en*.bin ggml-small.en*.bin ggml-large-v3-turbo-q8_0.bin note voice stack

huihui-ai/Huihui-Qwen-AgentWorld-35B-A3B-abliterated

include model-*.safetensors

huihui-ai/Huihui-Qwen3-Coder-Next-abliterated

include model-*.safetensors

jaredpalmer/kev-0.8b

note decision model

jaredpalmer/kev-4b

note decision model

jaredpalmer/kev-9b

note decision model

jlnsrk/GLM-5.2-colibri-int4

note colibri-format (Colibrì engine): GLM-5.2 744B int4, 379 GB, NVMe-streamed experts

Justvugg/GLM-5.3-colibri-int4-g64

note colibri-format (Colibrì engine, author's own checkpoint): GLM-5.3 744B int4-g64, 419 GB, NVMe-streamed experts

Kanposer/DeepSeek-V4-Flash-0731-speedy-colibri-mxfp4

note colibri-format (Colibrì engine): DeepSeek V4 Flash 0731 mxfp4, 155 GB

Krishal/QWen3.6-35B-A3B-Colibri

note colibri-format (Colibrì engine)

LocalLaws/LOCUS-Enforcement-Discretion

note LOCUS ModernBERT ordinance classifier/scorer (Enforcement-Discretion), pairs with LocalLaws/LOCUS-v1 dataset

LocalLaws/LOCUS-Function

note LOCUS ModernBERT ordinance classifier/scorer (Function), pairs with LocalLaws/LOCUS-v1 dataset

LocalLaws/LOCUS-Opacity

note LOCUS ModernBERT ordinance classifier/scorer (Opacity), pairs with LocalLaws/LOCUS-v1 dataset

LocalLaws/LOCUS-Paternalism

note LOCUS ModernBERT ordinance classifier/scorer (Paternalism), pairs with LocalLaws/LOCUS-v1 dataset

LocalLaws/LOCUS-Problem-Salience

note LOCUS ModernBERT ordinance classifier/scorer (Problem-Salience), pairs with LocalLaws/LOCUS-v1 dataset

LocalLaws/LOCUS-Substantive

note LOCUS ModernBERT ordinance classifier/scorer (Substantive), pairs with LocalLaws/LOCUS-v1 dataset

LocalLaws/LOCUS-Topic

note LOCUS ModernBERT ordinance classifier/scorer (Topic), pairs with LocalLaws/LOCUS-v1 dataset

Mapika/decider-0.8b

note decision model

Mapika/decider-2b

note decision model

Mapika/decider-2b-vision

note decision model

Mapika/decider-35b-a3b

note decision model

Mapika/decider-4b

note decision model

meta-models/Muse-Glimmer-30B

note agentic 30B VLM (Meta), bf16 safetensors

microsoft/Florence-2-large

note Florence-2 large (arXiv 2311.06242): OCR + region tasks; small fast baseline

microsoft/trocr-large-printed

note TrOCR large printed (arXiv 2109.10282): line-level transformer OCR baseline

nanonets/Nanonets-OCR2-3B

note Nanonets-OCR2 3B: markdown/semantic OCR (tables, equations, signatures)

NousResearch/Minos-v1

include model.safetensors *.json README.md examples/* note refusal classifier (ModernBERT, chat-pair input): detects model refusals for the Pod adaptive-refusal proxy

opendatalab/MinerU2.5-2509-1.2B

note MinerU2.5 1.2B (arXiv 2509.22186): two-stage document parsing (layout at low res, recognition at native res)

protectai/distilroberta-base-rejection-v1

include model.safetensors *.json merges.txt README.md onnx/* note refusal/rejection classifier (DistilRoBERTa, ONNX included): CPU-cheap refusal detection

Qwen/Qwen3-Coder-Next-FP8

include model-*.safetensors

Qwen/Qwen3-Embedding-0.6B

include model.safetensors

Qwen/Qwen3-Reranker-0.6B

include model.safetensors

Qwen/Qwen3-VL-32B-Instruct

include model-*.safetensors

Qwen/Qwen3.5-122B-A10B-FP8

include model.safetensors-*.safetensors

Qwen/Qwen3.5-4B-Base

note base for jaredpalmer/kev-4b (PEFT LoRA + head.pt load only on this exact snapshot; OC-11 open-chat eval)

Qwen/Qwen3.5-9B

include model.safetensors-*.safetensors

Qwen/Qwen3.8-Flash-Next

note Qwen3.8-Flash-Next raw release (335 GiB, 144 files, BF16 safetensors + n-gram embedding shards): source for strixite's convert_qwen4exp/convert_ngram_table (strixw) and for CPU goldens; fetched 2026-10-04 to bigboy:/mnt/scratch1t/ark-inbox/models

rednote-hilab/dots.ocr

note dots.ocr (1.7B): layout + OCR in one VLM, multilingual

stepfun-ai/GOT-OCR-2.0-hf

note GOT-OCR2.0 (580M, arXiv 2409.01704): general OCR theory model, HF transformers port

SupraLabs/SupraElegans-500k

note 500k-param sparse recurrent LM (no attention), BCI/biosignal-adjacent architecture experiment

Tongyi-MAI/Z-Image-Turbo

include text_encoder/model-*.safetensors transformer/diffusion_pytorch_model-*.safetensors vae/diffusion_pytorch_model.safetensors

Wan-AI/Wan2.2-S2V-14B

include Wan2.1_VAE.pth diffusion_pytorch_model-*.safetensors models_t5_umt5-xxl-enc-bf16.pth wav2vec2-large-xlsr-53-english/flax_model.msgpack wav2vec2-large-xlsr-53-english/language_model/lm.binary wav2vec2-

Wan-AI/Wan2.2-TI2V-5B

include Wan2.2_VAE.pth diffusion_pytorch_model-*.safetensors models_t5_umt5-xxl-enc-bf16.pth

wemoh/Qwen3.8-Flash-Next-strixw

note Qwen3.8-Flash-Next in strixw (strixite's own format; only strixite loads it): weights.strixw 65.9 GiB (4-bit experts, 8-bit dense, MTP) + ngram.table 48.3 GiB (int8, read from SSD on demand) + tokenizer/config + sha256.txt; ~115 GiB. uploaded 2026-10-05 (122.6 GB); place at ~/models/strix-infer, sha256sum -c sha256.txt

wfzyx/von

note decision model

zai-org/GLM-4.6V-FP8

include model-*.safetensors

zai-org/GLM-OCR

include model.safetensors

zai-org/GLM-TTS

include flow/flow.pt hift/hift.pt llm/model-*.safetensors speech_tokenizer/model.safetensors vocos2d/generator_jit.ckpt

zai-org/Glyph

note Glyph (Zhipu): GLM-4.1V fine-tuned to read rendered long text, 3-4x compression; general-VLM reasoner candidate for optical memory

zai-org/SCAIL-2

include Wan2.1_VAE.pth model/1/fsdp2_rank_0000_checkpoint.pt model/bias-aware-dpo-lora.pt model/relighting-lora.pt models_clip_open-clip-xlm-roberta-large-vit-huge-14-onlyvisual.pth umt5-xxl/models_t5_umt5-xx

zichenshang/Ornith-1.5-35B-colibri-i8

note colibri-format (Colibrì engine)

Model picks: GGUF quants (llama.cpp)32

AngelSlim/Hy4-preview-GGUF

include Hy4-preview-Q4_K_M.gguf

antirez/deepseek-v4-gguf

include DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-0731.gguf DeepSeek-V4-Flash-DSpark-support-0731.gguf DeepSeek-V4-Flash-MTP-Q4K-Q8_0-F32.gguf DeepSeek-V4-Flash-Vision-Encoder.gguf README.md note DeepSeek V4 Flash 0731, antirez IQ2XXS (81 GiB, fits ONE Strix Halo) + DSpark draft + MTP + vision encoder: gufo single-node 18-21 t/s on the Pod (2026-10-03), also the ds4 TP=2 input

bartowski/huihui-ai_Qwen3-Coder-Next-abliterated-GGUF

include huihui-ai_Qwen3-Coder-Next-abliterated-Q8_0/huihui-ai_Qwen3-Coder-Next-abliterated-Q8_0-*.gguf

crh225/plumb-4b-GGUF

note decision model

CrocoCpp/Huihui-Qwen-AgentWorld-35B-A3B-abliterated-Q8_0-GGUF

include huihui-qwen-agentworld-35b-a3b-abliterated-q8_0.gguf

ggml-org/embeddinggemma-300M-GGUF

include embeddinggemma-300M-Q8_0.gguf note EmbeddingGemma 300M Q8_0 GGUF (official ggml-org, only quant offered; llama-embedding/llama-server --embeddings): llama.cpp CPU/iGPU baseline for the FastFlowLM NPU2 embed-gemma (NP-2)

ggml-org/MiMo-V2.6-Pro-RL-GGUF

include MiMo-V2.6-Pro-RL-MXFP4/* mmproj-MiMo-V2.6-Pro-RL-Q8_0.gguf mtp-MiMo-V2.6-Pro-RL-Q8_0.gguf dflash-MiMo-V2.6-Pro-RL-Q8_0.gguf note MiMo-V2.6-Pro-RL MXFP4 (553 GB) + vision mmproj + MTP and DFlash speculative-decoding sidecars; replaces AesSedai/MiMo-V2.6-Pro-RL-GGUF (repo deleted Oct 2026), whose 13-way MXFP4 split of the same quant is what the reference island holds

ggml-org/Qwen3-ASR-1.7B-GGUF

include Qwen3-ASR-1.7B-Q8_0.gguf mmproj-Qwen3-ASR-1.7B-Q8_0.gguf

ggml-org/Qwen3-TTS-12Hz-1.7B-Base-GGUF

include Qwen3-TTS-12Hz-1.7B-Base-Q8_0.gguf mmproj-Qwen3-TTS-12Hz-1.7B-Base-Q8_0.gguf

hotdogs/Qwen35B-Agent-R2-Abliterated

include GGUF/Qwen35B-Agent-R2-Abliterated.Q6_K.gguf model-*.safetensors

hotdogs/Qwen35b-agent-R2O3

include GGUF-MTP/Qwen35b-agent-R2O3-MTP.Q6_K.gguf

ilintar/trellis2-gguf

include birefnet.gguf dinov3.gguf ss_flow.gguf ss_dec.gguf shape_flow_512.gguf shape_flow_1024.gguf shape_dec.gguf tex_flow_512.gguf tex_flow_1024.gguf tex_dec.gguf README.md note trellis.cpp weight set (full precision, 16.5 GB; q8/ and q4/ subsets also published); point --models at the folder

Mapika/decider-2b-GGUF

note decision model

Mapika/decider-4b-GGUF

note decision model

mys/laya-GGUF

note decision model

mys/laya-multilingual-GGUF

note decision model

mys/laya-typed-decisions-GGUF

note decision model

onnx-community/kev-0.6b-ONNX

note decision model

ornith-ai/Ornith-1.5-35B-A3B-GGUF

include Ornith-1.5-35B-Q8_0.gguf mmproj-Ornith-1.5-35B-BF16.gguf

ornith-ai/Ornith-1.5-397B-GGUF

include Ornith-1.5-397B-Q4_K_M.gguf mmproj-Ornith-1.5-397B-BF16.gguf

unsloth/DeepSeek-V4-Flash-0731-GGUF

include UD-Q8_K_XL dspark-DeepSeek-V4-Flash-0731-Q8_0.gguf note V4 Flash July-31 refresh: native MXFP4 experts + Q8 attention (151 GiB, fits two 128 GB nodes resident); the MindMeld two-node model. Forgotten until 2026-10-01; fetched to bigboy scratch1t then pod-control /data/models/deepseek-v4-flash-0731

unsloth/gemma-4-12B-it-qat-GGUF

include gemma-4-12B-it-qat-UD-Q4_K_XL.gguf mmproj-F16.gguf MTP/mtp-gemma-4-12B-it-Q8_0.gguf note halo's reference model: iGPU 329 prefill / ~26 decode tok/s on Strix Halo; MTP head for atomic-llama --spec-type mtp

unsloth/GLM-5.3-GGUF

include UD-Q4_K_XL/*

unsloth/Kimi-K2.7-Code-GGUF

include UD-Q3_K_XL mmproj-BF16.gguf

unsloth/MiniMax-M3-GGUF

include UD-Q4_K_XL/*

unsloth/Mistral-Medium-3.5-128B-GGUF

include UD-Q4_K_XL/*

unsloth/Muse-Glimmer-30B-GGUF

include Muse-Glimmer-30B-Q8_0.gguf dflash-kquant.gguf mmproj-Muse-Glimmer-30B-Q8_0.gguf

unsloth/NVIDIA-Nemotron-3-Ultra-550B-A55B-GGUF

include UD-Q4_K_XL/*

unsloth/Qwen3-4B-Instruct-2507-GGUF

include Qwen3-4B-Instruct-2507-Q4_K_M.gguf note llama.cpp iGPU/CPU baseline for FastFlowLM qwen3-it:4b on the NPU (NP-2 job 4)

unsloth/Qwen3.8-27B-GGUF

include MTP/mtp-Qwen3.8-27B-Q4_0.gguf Qwen3.8-27B-Q8_0.gguf mmproj-BF16.gguf

unsloth/Qwen3.8-Flash-Next-GGUF

include MTP/mtp-Qwen3.8-Flash-Next-Q8_0.gguf MTP/mtp-Qwen3.8-Flash-Next-shared-Q8_0.gguf UD-IQ4_XS mmproj-BF16.gguf

vcruz305/DeepSeek-V4.1-Flash-GGUF

include DeepSeek-V4.1-Flash-Q4_K_M-*.gguf llama.cpp/patches/fix_gguf_engram_kv.py llama.cpp/patches/patch_llamacpp_v41.py

Dependencies: helper weights that mirrored code downloads at first run (not picks)63

aoi-ot/VibeVoice-Large

note VibeVoice-Large TTS (MIT): byte copy of microsoft/VibeVoice-Large from ModelScope; Microsoft pulled the HF repo (Sept 2025), the code is mirrored as github.com/microsoft/VibeVoice

canopylabs/orpheus-3b-0.1-ft

note voice stack

csukuangfj/sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8

note voice stack

cvssp/audioldm2

note audio tooling

cvssp/audioldm2-large

note audio tooling

declare-lab/tango2

note audio tooling

deepdml/faster-whisper-large-v3-turbo-ct2

note voice stack

diffusers/stable-diffusion-xl-1.0-inpainting-0.1

include *.json *.txt *.fp16.safetensors note SDXL 9-channel inpainting UNet (fp16 variant); ImageBoop concept-injection fixer (inject/fixer.py)

facebook/audiogen-medium

note audio tooling

facebook/wav2vec2-base-960h

note audio tooling

fishaudio/s1-mini

note voice stack

fixie-ai/ultravox-v0_5-llama-3_1-8b

note voice stack

ggml-org/Qwen3-Omni-30B-A3B-Instruct-GGUF

include *Q8_0.gguf mmproj-* README.md note voice stack

hexgrad/Kokoro-82M

note voice stack

jonatasgrosman/wav2vec2-large-xlsr-53-english

note audio tooling

Kijai/DepthAnythingV2-safetensors

note ComfyUI-DepthAnythingV2 node weights

Kijai/GIMM-VFI_safetensors

note ComfyUI-GIMM-VFI node weights

Kijai/HunyuanVideo_comfy

include hunyuan_video_720_cfgdistill_fp8_e4m3fn.safetensors hunyuan_video_I2V_720_fixed_fp8_e4m3fn.safetensors hunyuan_video_FastVideo_720_fp8_e4m3fn.safetensors FramePackI2V_HY_fp8_e4m3fn.safetensors hunyuan_video_vae_*.safetensors hy*lora*.safetensors note collection is 300+ GB unfiltered; keep the loadable fp8 set

Kijai/LivePortrait_safetensors

note ComfyUI-LivePortraitKJ node weights

Kijai/MMAudio_safetensors

note comfy node weights

Kijai/sam2-safetensors

note ComfyUI-segment-anything-2 node weights

Kijai/SUPIR_pruned

note ComfyUI-SUPIR node weights

kyutai/moshiko-pytorch-bf16

note voice stack

kyutai/pocket-tts

note voice stack

kyutai/stt-1b-en_fr

note voice stack

kyutai/stt-2.6b-en

note voice stack

kyutai/tts-1.6b-en_fr

note voice stack

laion/CLIP-ViT-H-14-laion2B-s32B-b79K

include model.safetensors config.json preprocessor_config.json tokenizer.json tokenizer_config.json vocab.json merges.txt special_tokens_map.json open_clip_config.json README.md note OpenCLIP ViT-H/14 laion2B as a transformers CLIPModel (safetensors only; skips the 3 duplicate 3.9 GB checkpoints): the CLIP_TEXT_REPO the afference engine expects for its text-defined affect axes (AE-8; lane substituted a merged SDXL bigG stand-in 2026-10-04 because this was missing)

LBH-123-AI/Minimax_h3_latent_Upscaler

include minimax_h3_latent_upscaler_3d_conv_v1/minimax_h3_latent_upscaler_3d_conv_v1_fp16.safetensors minimax_h3_latent_upscaler_3d_conv_v1/config.json README.md note H3 learned 3D latent upscaler (h3 --upscale)

livekit/turn-detector

note voice stack

lj1995/VoiceConversionWebUI

include hubert_base.pt rmvpe.pt rmvpe.onnx pretrained_v2/* uvr5_weights/HP2-* uvr5_weights/HP5-* uvr5_weights/VR-DeEchoAggressive.pth RVC20260718Nvidia50x0.7z README.md note RVC voice conversion: hubert + rmvpe + v2 pretrains + three UVR5 separators + the latest release bundle (9.2 GB of a 103 GB repo of legacy files)

lllyasviel/fooocus_inpaint

note comfyui-inpaint-nodes weights

microsoft/TRELLIS-image-large

note TRELLIS node weights

microsoft/TRELLIS.2-4B

note TRELLIS.2 source checkpoints (ComfyUI-Trellis2 node, trellis.cpp tools/convert.py)

microsoft/VibeVoice-1.5B

note audio tooling

mistralai/Voxtral-4B-TTS-2603

note voice stack

mistralai/Voxtral-Mini-3B-2507

note voice stack

mistralai/Voxtral-Mini-4B-Realtime-2602

note voice stack

moonshine-ai/moonshine-base

note voice stack

nari-labs/Dia-1.6B-0626

note voice stack

nvidia/parakeet-tdt-0.6b-v3

note voice stack

openai/whisper-large-v3-turbo

note voice stack

peonist-ai/halogen-qwen3.8-flash-next

include qwen38-flash-next-mtp.hgn qwen38-flash-next-vision.hgn tokenizer/* note draft head (1.5 GB), vision tower (0.9 GB) and tokenizer that halogen-flash-server needs beside our unsloth GGUF; the full .hgn checkpoints (v2 67 GB + ngram 51 GB, 5.53 bpw) are optional and only load in halogen

pipecat-ai/smart-turn-v3

note voice stack

Plachta/Seed-VC

note audio tooling

Qwen/Qwen3-Omni-30B-A3B-Instruct

note voice stack

Qwen/Qwen3-TTS-12Hz-0.6B-Base

note Qwen3-TTS 0.6B Base (2.5 GB, Apache-2.0): the model navi-tts runs (convert once with navi-tts tools/convert.py -> .navi)

ResembleAI/chatterbox

note voice stack

ResembleAI/resemble-enhance

note audio tooling

rhasspy/piper-voices

include en/en_US/* en/en_GB/* voices.json note voice stack

sesame/csm-1b

note voice stack

stabilityai/stable-audio-open-1.0

note audio tooling

stabilityai/stable-audio-open-small

note audio tooling

stanford-crfm/music-large-800k

note audio tooling

stanford-crfm/music-medium-800k

note audio tooling

stepfun-ai/Step-Audio-2-mini

note voice stack

tencent/Hunyuan3D-2.1

note ComfyUI-Hunyuan3DWrapper weights

tianweiy/DMD2

include dmd2_sdxl_4step_unet_fp16.bin dmd2_sdxl_1step_unet_fp16.bin dmd2_sdxl_4step_lora_fp16.safetensors model/sdv1.5/laion6.25_sd_baseline_8node_guidance1.75_lr5e-7_seed10_dfake10_diffusion1000_gan1e-3_resume_fid8.35_checkpoint_model_041000/pytorch_model.bin note DMD2 distilled SDXL (4/1-step UNet, 4-step LoRA) + SD1.5 1-step generator checkpoint (no SD1.5 4-step exists upstream); ImageBoop/Pareidolia fast renders

unsloth/embeddinggemma-300m

include *.safetensors *.json tokenizer.model README.md note EmbeddingGemma 300M (sentence-transformers layout): ungated copy of google/embeddinggemma-300m; CPU/iGPU baseline for the FastFlowLM NPU2 embed-gemma (NP-2)

zai-org/glm-4-voice-9b

note voice stack

zai-org/glm-4-voice-decoder

note voice stack

zai-org/glm-4-voice-tokenizer

note voice stack

ZhengPeng7/BiRefNet

note BiRefNet matting (ComfyUI_BiRefNet_ll, rembg)

Dependencies: ComfyUI assets (checkpoints, LoRAs, VAEs, ControlNets)29

black-forest-labs/FLUX.2-klein-9b-kv-fp8

include flux-2-klein-9b-kv-fp8.safetensors note ComfyUI asset (checkpoint/lora/vae/controlnet)

Comfy-Org/ace_step_1.5_ComfyUI_files

include checkpoints/ace_step_1.5_turbo_aio.safetensors note ComfyUI asset (checkpoint/lora/vae/controlnet)

Comfy-Org/BiRefNet

include background_removal/birefnet.safetensors note ComfyUI built-in background-removal node (TRELLIS.2 / Pixal3D workflows), models/background_removal/

Comfy-Org/Chroma1-HD_repackaged

include split_files/diffusion_models/Chroma1-HD-fp8mixed.safetensors note ComfyUI asset (checkpoint/lora/vae/controlnet)

Comfy-Org/flux1-dev

include flux1-dev-fp8.safetensors note ComfyUI asset (checkpoint/lora/vae/controlnet)

Comfy-Org/flux1-kontext-dev_ComfyUI

include split_files/diffusion_models/flux1-dev-kontext_fp8_scaled.safetensors note ComfyUI asset (checkpoint/lora/vae/controlnet)

Comfy-Org/flux1-schnell

include flux1-schnell-fp8.safetensors note ComfyUI asset (checkpoint/lora/vae/controlnet)

Comfy-Org/flux2-dev

include split_files/diffusion_models/flux2_dev_fp8mixed.safetensors split_files/text_encoders/mistral_3_small_flux2_fp8.safetensors note ComfyUI asset (checkpoint/lora/vae/controlnet)

Comfy-Org/gemma-4

include text_encoders/gemma4_e4b_it_fp8_scaled.safetensors note ComfyUI asset (checkpoint/lora/vae/controlnet)

Comfy-Org/HiDream-O1-Image

include checkpoints/hidream_o1_image_dev_fp8_scaled.safetensors note ComfyUI asset (checkpoint/lora/vae/controlnet)

Comfy-Org/Krea-2

include diffusion_models/krea2_raw_fp8_scaled.safetensors diffusion_models/krea2_turbo_fp8_scaled.safetensors note ComfyUI asset (checkpoint/lora/vae/controlnet)

Comfy-Org/ltx-2

include split_files/text_encoders/gemma_3_12B_it.safetensors note ComfyUI asset: Gemma 3 12B text encoder for LTX-2.3 (DarkStudio render.py LTX_TE)

Comfy-Org/MiniMax-H3

include diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors text_encoders/qwen3vl_32b_minimax_h3_int8_convrot.safetensors vae/minimax_h3_video_vae_fp16.safetensors vae/minimax_h3_audio_vae_fp32.safetensors note the five files h3-hrx loads (int8_convrot DiTs 2x21 GB; TE+VAEs also in nv6 comfy/packages/minimax-h3)

Comfy-Org/Pixal3D

include clip_vision/dino_v3_L_naf_fp32.safetensors diffusion_models/pixal3d_bf16.safetensors diffusion_models/pixal3d_multiview_bf16.safetensors vae/trellis_2_shape_vae_bf16.safetensors vae/trellis_2_texture_ note ComfyUI asset (checkpoint/lora/vae/controlnet)

Comfy-Org/Qwen-Image-2.1

include diffusion_models/qwen_image_2.1_bf16.safetensors vae/qwen_image_2.1_vae_bf16.safetensors note ComfyUI asset (checkpoint/lora/vae/controlnet)

Comfy-Org/Qwen3-VL

include text_encoders/qwen3vl_4b_fp8_scaled.safetensors text_encoders/qwen3vl_8b_fp8_scaled.safetensors note ComfyUI asset (checkpoint/lora/vae/controlnet)

Comfy-Org/stable-diffusion-v1-5-archive

include v1-5-pruned-emaonly-fp16.safetensors note ComfyUI asset (checkpoint/lora/vae/controlnet)

Comfy-Org/TRELLIS.2

include clip_vision/dino_v3_vit_l.safetensors diffusion_models/trellis_2_bf16.safetensors note ComfyUI asset (checkpoint/lora/vae/controlnet)

Comfy-Org/vae-text-encorder-for-flux-klein-9b

include split_files/text_encoders/qwen_3_8b_fp8mixed.safetensors split_files/vae/flux2-vae.safetensors note ComfyUI asset (checkpoint/lora/vae/controlnet)

Comfy-Org/Wan_2.1_ComfyUI_repackaged

include split_files/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors note ComfyUI asset: umt5-xxl fp8 text encoder for Wan 2.1/2.2 (DarkStudio render.py wan_te)

Comfy-Org/z_image

include split_files/diffusion_models/z_image_bf16.safetensors split_files/text_encoders/qwen_3_4b_fp8_mixed.safetensors split_files/vae/ae.safetensors note ComfyUI asset (checkpoint/lora/vae/controlnet)

comfyanonymous/flux_text_encoders

include clip_l.safetensors t5xxl_fp8_e4m3fn_scaled.safetensors note ComfyUI asset (checkpoint/lora/vae/controlnet)

Lightricks/LTX-2.3

include ltx-2.3-22b-distilled-1.1.safetensors note ComfyUI asset: LTX-2.3 distilled 1.1 checkpoint (DarkStudio render.py LTX_CKPT; video+audio, 8-step)

Lightricks/LTX-2.5

include diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors vae/ltx-2.5-audio-vae-bf16.safetensors vae/ltx-2.5-video-vae-bf16.safete note ComfyUI asset (checkpoint/lora/vae/controlnet)

OnomaAIResearch/Illustrious-xl-early-release-v0

include Illustrious-XL-v0.1.safetensors note ComfyUI asset (checkpoint/lora/vae/controlnet)

OnomaAIResearch/Illustrious-XL-v2.0

include Illustrious-XL-v2.0.safetensors note ComfyUI asset (checkpoint/lora/vae/controlnet)

PixArt-alpha/PixArt-Sigma-XL-2-1024-MS

include transformer/diffusion_pytorch_model.safetensors vae/diffusion_pytorch_model.safetensors note ComfyUI asset (checkpoint/lora/vae/controlnet)

realrebelai/Wan-Animate-2_GGUFs

include Distilled/Wan-Animate-2-14B-TURBO-Q8_0.gguf note ComfyUI asset (checkpoint/lora/vae/controlnet)

Viggle/Qwen-Image-2.1-viggle-turbo

include Qwen-Image-2.1-viggle-turbo-4step-lora-r64.safetensors Qwen-Image-2.1-viggle-turbo-v0.2.1-6step-lora-r128.safetensors note ComfyUI asset (checkpoint/lora/vae/controlnet)

Datasets29

allenai/tulu-3-sft-mixture

note SFT

Anonymous112233/FlowGen

note FlowGen: synthetic flowcharts per renderer (Mermaid/Graphviz/PlantUML/Diagrams), 3840 train / 1440 test each, by difficulty

caiovicentino1/ekbasis-data

note Ekbasis evaluation + training data (0.44 GB): git3 known/held, ftest_family, multi_test sets used by its pre-registered quantized-build gate

cais/mmlu

note eval

glaiveai/glaive-code-assistant-v3

note code instruction (Apache-2.0)

HuggingFaceFW/fineweb-edu

include sample/10BT/* note pretraining sample (10B tokens)

HuggingFaceH4/MATH-500

note eval

HuggingFaceH4/ultrafeedback_binarized

note preference/DPO

HuggingFaceTB/smoltalk

note SFT

ise-uiuc/Magicoder-OSS-Instruct-75K

note code instruction (MIT)

jaredpalmer/kev-suites

note decision-model eval

jopan/FlowLearn

note FlowLearn: scientific + simulated flowcharts with Mermaid, OCR and VQA annotations

keithito/lj_speech

note TTS training

linoyts/Tuxemon

note image LoRA (tiny)

LocalLaws/LOCUS-Fines

note LOCUS fines/penalties supplement (87 MB parquet)

LocalLaws/LOCUS-v1

note LOCUS v1 (Berkeley Responsibility Lab Local Law Observatory): 2.2M sections of US municipal+county ordinance law, 2309 counties, labelled function/topic/opacity/paternalism/discretion/salience (1.8 GB parquet, CC-BY-NC, arXiv 2606.19334)

MiaSanLei/MindBench

note MindBench synthetic mind maps (Graphviz, bilingual) with structure annotations; CC BY-NC 4.0| full set, dad 2026-10-13 (DECIDE a4141f7018): pull everything, straight to /mnt/ark/datasets via tools/dataset-courier.sh (never through /data/ark-inbox)

MLLM-CL/VTCBench

note VTCBench: long-context retrieval/reasoning/memory under vision-text compression (OMF H2 public comparator)

MrSunshy/VisionGraph_GraphStructure_hf

note VisionGraph: rendered graphs + node/edge perception and graph-theory QA| full set, dad 2026-10-13 (DECIDE a4141f7018): pull everything, straight to /mnt/ark/datasets via tools/dataset-courier.sh (never through /data/ark-inbox)

nlphuji/flickr30k

note Flickr30k captions (was on the crucial drive)

nvidia/HelpSteer3

note preference

nvidia/OpenCodeInstruct

note code instruction, large (CC-BY-4.0)

open-r1/OpenR1-Math-220k

note reasoning SFT

openai/gsm8k

note eval

openai/openai_humaneval

note eval (code)

pcuenq/oxford-pets

note image LoRA

Rowan/hellaswag

note eval

tanganke/stanford_cars

note Stanford Cars (was on the crucial drive)

teknium/OpenHermes-2.5

note SFT

Reference data146

Kiwix ZIMs (latest match of a pattern wins), DevDocs sets, GRCh38 genome references with {{REL}} resolved to the current Ensembl release.

Offline knowledge (Kiwix ZIM)4

Wikipedia (English, no pictures), latest; ~53 GBzim · wikipedia_en_all_nopic_*.zim

source https://download.kiwix.org/zim/wikipedia/

Stack Overflow, latest; ~115 GBzim · stackoverflow.com_en_all_*.zim

source https://download.kiwix.org/zim/stack_exchange/

Project Gutenberg (English), latest; ~221 GBzim · gutenberg_en_all_*.zim

source https://download.kiwix.org/zim/gutenberg/

DevDocs: latest ZIM per listed language/tool (~70 sets)zim-set · devdocs_en_{python,pytorch,numpy,pandas,matplotlib,scikit_learn,javascript,typescript,node,npm,react,vue,svelte,vite,webpack,jest,vitest,eslint,threejs,d3,electron,tailwindcss,html,css,http,markdown,rust,c,cpp,gcc,cmake,go,zig,lua,julia,r,kotlin,scala,haskell,elixir,erlang,php,ruby,bash,git,docker,kubernetes,terraform,ansible,nix,nginx,redis,postgresql,sqlite,fastapi,flask,django,latex,godot,opengl,vulkan,qt,gtk,jq}_*.zim

source https://download.kiwix.org/zim/devdocs/

Genome reference data (GRCh38)7

GRCh38 fasta (UCSC)url · hg38.fa.gz

source https://hgdownload.soe.ucsc.edu/goldenPath/hg38/bigZips/hg38.fa.gz lands in genome

Ensembl primary assemblyurl · Homo_sapiens.GRCh38.dna.primary_assembly.fa.gz

source https://ftp.ensembl.org/pub/release-{REL}/fasta/homo_sapiens/dna/Homo_sapiens.GRCh38.dna.primary_assembly.fa.gz lands in genome

Gene annotation GTF (Ensembl)url · Homo_sapiens.GRCh38.<release>.gtf.gz

source https://ftp.ensembl.org/pub/release-{REL}/gtf/homo_sapiens/Homo_sapiens.GRCh38.{REL}.gtf.gz lands in genome

VEP cache (~28 GB)url · homo_sapiens_vep_<release>_GRCh38.tar.gz

source https://ftp.ensembl.org/pub/release-{REL}/variation/indexed_vep_cache/homo_sapiens_vep_{REL}_GRCh38.tar.gz lands in genome

ClinVar (+ .tbi)url · clinvar.vcf.gz

source https://ftp.ncbi.nlm.nih.gov/pub/clinvar/vcf_GRCh38/clinvar.vcf.gz lands in genome

dbSNP GRCh38.p14 (~30 GB, + .tbi)url · GCF_000001405.40.gz

source https://ftp.ncbi.nih.gov/snp/latest_release/VCF/GCF_000001405.40.gz lands in genome

gnomAD v4.1 exomes, chr22 sample onlyurl · gnomad.exomes.v4.1.sites.chr22.vcf.bgz

source https://storage.googleapis.com/gcp-public-data--gnomad/release/4.1/vcf/exomes/gnomad.exomes.v4.1.sites.chr22.vcf.bgz lands in genome

Paper shelf (Leitner Reference Library)49

FlashAttentionurl · 2205.14135

source https://arxiv.org/pdf/2205.14135 lands in papers

FlashAttention-2url · 2307.08691

source https://arxiv.org/pdf/2307.08691 lands in papers

FlashAttention-3url · 2407.08608

source https://arxiv.org/pdf/2407.08608 lands in papers

FlashDecoding++url · 2311.01282

source https://arxiv.org/pdf/2311.01282 lands in papers

speculative decoding (Leviathan)url · 2211.17192

source https://arxiv.org/pdf/2211.17192 lands in papers

EAGLE-2url · 2406.16858

source https://arxiv.org/pdf/2406.16858 lands in papers

Medusaurl · 2401.10774

source https://arxiv.org/pdf/2401.10774 lands in papers

multi-token predictionurl · 2404.19737

source https://arxiv.org/pdf/2404.19737 lands in papers

GPTQurl · 2210.17323

source https://arxiv.org/pdf/2210.17323 lands in papers

AWQurl · 2306.00978

source https://arxiv.org/pdf/2306.00978 lands in papers

SmoothQuanturl · 2211.10438

source https://arxiv.org/pdf/2211.10438 lands in papers

OCP Microscaling (MX) paper; the spec itself is registration-gatedurl · 2310.10537

source https://arxiv.org/pdf/2310.10537 lands in papers

LoRAurl · 2106.09685

source https://arxiv.org/pdf/2106.09685 lands in papers

DoRAurl · 2402.09353

source https://arxiv.org/pdf/2402.09353 lands in papers

flow matchingurl · 2210.02747

source https://arxiv.org/pdf/2210.02747 lands in papers

rectified flowurl · 2209.03003

source https://arxiv.org/pdf/2209.03003 lands in papers

GShardurl · 2006.16668

source https://arxiv.org/pdf/2006.16668 lands in papers

Switch Transformerurl · 2101.03961

source https://arxiv.org/pdf/2101.03961 lands in papers

Megatron-LMurl · 1909.08053

source https://arxiv.org/pdf/1909.08053 lands in papers

MegaBlocksurl · 2211.15841

source https://arxiv.org/pdf/2211.15841 lands in papers

efficiently scaling transformer inference (Pope)url · 2211.05102

source https://arxiv.org/pdf/2211.05102 lands in papers

data movement is all you needurl · 2007.00072

source https://arxiv.org/pdf/2007.00072 lands in papers

DeepSeek-V2 (MLA)url · 2405.04434

source https://arxiv.org/pdf/2405.04434 lands in papers

DeepSeek-V3 reporturl · 2412.19437

source https://arxiv.org/pdf/2412.19437 lands in papers

DeepSeek-V4 reporturl · 2606.19348

source https://arxiv.org/pdf/2606.19348 lands in papers

FlashMemory-DeepSeek-V4 (lookahead sparse attention)url · 2606.09079

source https://arxiv.org/pdf/2606.09079 lands in papers

LTX-Videourl · 2501.00103

source https://arxiv.org/pdf/2501.00103 lands in papers

LTX-2 reporturl · 2601.03233

source https://arxiv.org/pdf/2601.03233 lands in papers

Qwen3-VL reporturl · 2511.21631

source https://arxiv.org/pdf/2511.21631 lands in papers

DeepSeek-OCR: Contexts Optical Compression (arXiv 2510.18234)url · 2510.18234

source https://arxiv.org/pdf/2510.18234 lands in papers/optical-memory

DeepSeek-OCR 2: Visual Causal Flow (arXiv 2601.20552)url · 2601.20552

source https://arxiv.org/pdf/2601.20552 lands in papers/optical-memory

Unlimited OCR Works (Baidu, R-SWA constant KV, dozens of pages per 32K pass) (arXiv 2606.23050)url · 2606.23050

source https://arxiv.org/pdf/2606.23050 lands in papers/optical-memory

Glyph: scaling context windows via visual-text compression (THU CoAI + Zhipu) (arXiv 2510.17800)url · 2510.17800

source https://arxiv.org/pdf/2510.17800 lands in papers/optical-memory

Text or Pixels? It Takes Half: token efficiency of visual text inputs in multimodal LLMs (arXiv 2510.18279)url · 2510.18279

source https://arxiv.org/pdf/2510.18279 lands in papers/optical-memory

Vision-centric token compression in large language models (arXiv 2502.00791)url · 2502.00791

source https://arxiv.org/pdf/2502.00791 lands in papers/optical-memory

VisionZip: longer is better but not necessary in VLMs (visual token pruning) (arXiv 2412.04467)url · 2412.04467

source https://arxiv.org/pdf/2412.04467 lands in papers/optical-memory

GOT-OCR2.0: general OCR theory (arXiv 2409.01704)url · 2409.01704

source https://arxiv.org/pdf/2409.01704 lands in papers/optical-memory

olmOCR: unlocking trillions of tokens in PDFs with VLMs (arXiv 2502.18443)url · 2502.18443

source https://arxiv.org/pdf/2502.18443 lands in papers/optical-memory

MinerU2.5: decoupled two-stage document parsing (arXiv 2509.22186)url · 2509.22186

source https://arxiv.org/pdf/2509.22186 lands in papers/optical-memory

PaddleOCR-VL: 0.9B ultra-compact VLM for document parsing (arXiv 2510.14528)url · 2510.14528

source https://arxiv.org/pdf/2510.14528 lands in papers/optical-memory

Nougat: neural optical understanding for academic documents (arXiv 2308.13418)url · 2308.13418

source https://arxiv.org/pdf/2308.13418 lands in papers/optical-memory

Donut: OCR-free document understanding transformer (arXiv 2111.15664)url · 2111.15664

source https://arxiv.org/pdf/2111.15664 lands in papers/optical-memory

Pix2Struct: screenshot parsing as pretraining for visual language understanding (arXiv 2210.03347)url · 2210.03347

source https://arxiv.org/pdf/2210.03347 lands in papers/optical-memory

Kosmos-2.5: a multimodal literate model (arXiv 2309.11419)url · 2309.11419

source https://arxiv.org/pdf/2309.11419 lands in papers/optical-memory

TrOCR: transformer-based OCR with pretrained models (arXiv 2109.10282)url · 2109.10282

source https://arxiv.org/pdf/2109.10282 lands in papers/optical-memory

Florence-2: a unified representation for vision tasks (arXiv 2311.06242)url · 2311.06242

source https://arxiv.org/pdf/2311.06242 lands in papers/optical-memory

LayoutLMv3: pre-training for document AI with unified text and image masking (arXiv 2204.08387)url · 2204.08387

source https://arxiv.org/pdf/2204.08387 lands in papers/optical-memory

Qwen2.5-VL technical report (dynamic resolution, mRoPE, document parsing) (arXiv 2502.13923)url · 2502.13923

source https://arxiv.org/pdf/2502.13923 lands in papers/optical-memory

SigLIP 2: multilingual vision-language encoders (arXiv 2502.14786)url · 2502.14786

source https://arxiv.org/pdf/2502.14786 lands in papers/optical-memory

Hardware manuals and references3

Agner Fog instruction tablesurl · instruction_tables.pdf

source https://www.agner.org/optimize/instruction_tables.pdf lands in cpu

Agner Fog microarchitectureurl · microarchitecture.pdf

source https://www.agner.org/optimize/microarchitecture.pdf lands in cpu

Agner Fog optimizing assemblyurl · optimizing_assembly.pdf

source https://www.agner.org/optimize/optimizing_assembly.pdf lands in cpu

OS installer images1

Ubuntu 24.04 LTS live-server installer, latest point release + SHA256SUMS; the base of every node image and of PXE (casper kernel/initrd + squashfs served over HTTP)iso · ubuntu-24.04.*-live-server-amd64.iso

source https://releases.ubuntu.com/24.04/

Other reference files82

AMD PPR 57930 rev 3.00, Family 1Ah Model 70h A0 = Strix Halo (Ryzen AI Max+ 395, the Bosgame M5 nodes): core/MCA, APML, SB-TSI, DXIO, FCH registers; zip (served as 57930-A0-PUB_3.00.zip), sha256 6eb92cbc…0c52url · content

source https://docs.amd.com/api/khub/documents/fEY1GkRkcoEhenFftZyr8Q/content lands in hardware-docs

hellas.ai "thunderbolt-ibverbs: We have InfiniBand at home" (2026): RDMA over USB4 between Strix Halos, ~48 Gbit/s per direction, ~7 us one-way; index page (HTML only, no images)url ·

source https://blog.hellas.ai/blog/thunderbolt-ibverbs/ lands in rdma-clustering/hellas-blog/0-index

hellas.ai thunderbolt-ibverbs blog, chapter 1-background: background: why two-box inferenceurl ·

source https://blog.hellas.ai/blog/thunderbolt-ibverbs/1-background/ lands in rdma-clustering/hellas-blog/1-background

hellas.ai thunderbolt-ibverbs blog, chapter 2-scaling: scaling/networking options: Apple, GB10, Strix Halourl ·

source https://blog.hellas.ai/blog/thunderbolt-ibverbs/2-scaling/ lands in rdma-clustering/hellas-blog/2-scaling

hellas.ai thunderbolt-ibverbs blog, chapter 3-agentic-harness: how the driver was built with an agent harnessurl ·

source https://blog.hellas.ai/blog/thunderbolt-ibverbs/3-agentic-harness/ lands in rdma-clustering/hellas-blog/3-agentic-harness

hellas.ai thunderbolt-ibverbs blog, chapter 4-thunderbolt-ibverbs: the driver: design, perftest and vLLM/FSDP benchmarksurl ·

source https://blog.hellas.ai/blog/thunderbolt-ibverbs/4-thunderbolt-ibverbs/ lands in rdma-clustering/hellas-blog/4-thunderbolt-ibverbs

hellas.ai thunderbolt-ibverbs blog, chapter 5-closing: closing: MiniMax-M2.7 TP=2 ~20 tok/s, ~30% over TCP-over-Thunderbolturl ·

source https://blog.hellas.ai/blog/thunderbolt-ibverbs/5-closing/ lands in rdma-clustering/hellas-blog/5-closing

strixhalo.wiki: Clustering with RDMA (Oculink + ConnectX-3 InfiniBand on Strix Halo, ~26 Gbit/s PCIe 3.0 x4 cap, 1.12 us; links to the AMD RDMA cluster guide and ds4-vLLM)url · Clustering_with_RDMA

source https://strixhalo.wiki/AI/Clustering_with_RDMA lands in rdma-clustering/strixhalo-wiki

Level1Techs: "Ryzen AI Halo - USB4 Clustering with RDMA - Testing Notes", posts 1-20 (32 posts total)url · 253117

source https://forum.level1techs.com/t/ryzen-ai-halo-usb4-clustering-with-rdma-testing-notes/253117 lands in rdma-clustering/level1techs/page-1

Level1Techs USB4 RDMA testing notes, posts 21-32url · 253117?page=2

source https://forum.level1techs.com/t/ryzen-ai-halo-usb4-clustering-with-rdma-testing-notes/253117?page=2 lands in rdma-clustering/level1techs/page-2

Overmind full documentation as one text file (llms-full.txt, ~220 KB): capabilities graph, trace scoring, datasets, eval, Jev decision engine, optimisers, training, inference, MCP, CLI, self-hostingurl · llms-full.txt

source https://docs.overmindlab.ai/llms-full.txt lands in agent-platforms/overmind-docs

TheRock ROCm 10.1 nightly for gfx1151 (1.8 GiB), nearest surviving build to the 10.1.0a20260822 strixite was tested with; AMD keeps ~6 weeks of nightlies, so this URL WILL 404 later: the mirrored copy is the referenceurl · therock-dist-linux-gfx1151-10.1.0a20260823.tar.gz

source https://nightly.repo.amd.com/rocm/core/tarball/therock-dist-linux-gfx1151-10.1.0a20260823.tar.gz lands in toolchains/therock-gfx1151

TheRock ROCm 10.2 nightly for gfx1151 (1.9 GiB), newest on 2026-10-04; ages out like the one aboveurl · therock-dist-linux-gfx1151-10.2.0a20261004.tar.gz

source https://nightly.repo.amd.com/rocm/core/tarball/therock-dist-linux-gfx1151-10.2.0a20261004.tar.gz lands in toolchains/therock-gfx1151

trellis.cpp FetchContent: libwebp 1.5.0 (GLB WebP textures, default ON; offline: -DFETCHCONTENT_SOURCE_DIR_LIBWEBP=<extracted dir> or -DTRELLIS_WEBP=OFF)url · v1.5.0.tar.gz

source https://github.com/webmproject/libwebp/archive/refs/tags/v1.5.0.tar.gz lands in toolchains/trellis.cpp-fetchcontent

trellis.cpp optional retopo (TRELLIS_RETOPO_MATCHING): LEMON 1.3.1, URL_HASH-pinnedurl · lemon-1.3.1.tar.gz

source https://lemon.cs.elte.hu/pub/sources/lemon-1.3.1.tar.gz lands in toolchains/trellis.cpp-fetchcontent

trellis.cpp optional retopo (TRELLIS_RETOPO_COLLISION): CGAL 6.1.1, URL_HASH-pinnedurl · CGAL-6.1.1-library.tar.xz

source https://github.com/CGAL/cgal/releases/download/v6.1.1/CGAL-6.1.1-library.tar.xz lands in toolchains/trellis.cpp-fetchcontent

trellis.cpp optional retopo (TRELLIS_RETOPO_COLLISION): Boost 1.90.0, URL_HASH-pinnedurl · boost_1_90_0.tar.gz

source https://archives.boost.io/release/1.90.0/source/boost_1_90_0.tar.gz lands in toolchains/trellis.cpp-fetchcontent

trellis.cpp optional retopo (TRELLIS_RETOPO_COLLISION, Linux): GMP 6.3.0, URL_HASH-pinnedurl · gmp-6.3.0.tar.xz

source https://ftp.gnu.org/gnu/gmp/gmp-6.3.0.tar.xz lands in toolchains/trellis.cpp-fetchcontent

trellis.cpp optional retopo (TRELLIS_RETOPO_COLLISION, Linux): MPFR 4.2.2, URL_HASH-pinnedurl · mpfr-4.2.2.tar.xz

source https://www.mpfr.org/mpfr-4.2.2/mpfr-4.2.2.tar.xz lands in toolchains/trellis.cpp-fetchcontent

openWakeWord pre-trained ONNX model (the wheel downloads these at first run; offline needs them on disk)url · melspectrogram.onnx

source https://github.com/dscripka/openWakeWord/releases/download/v0.5.1/melspectrogram.onnx lands in voice/openwakeword-v0.5.1

openWakeWord pre-trained ONNX model (the wheel downloads these at first run; offline needs them on disk)url · embedding_model.onnx

source https://github.com/dscripka/openWakeWord/releases/download/v0.5.1/embedding_model.onnx lands in voice/openwakeword-v0.5.1

openWakeWord pre-trained ONNX model (the wheel downloads these at first run; offline needs them on disk)url · silero_vad.onnx

source https://github.com/dscripka/openWakeWord/releases/download/v0.5.1/silero_vad.onnx lands in voice/openwakeword-v0.5.1

openWakeWord pre-trained ONNX model (the wheel downloads these at first run; offline needs them on disk)url · hey_jarvis_v0.1.onnx

source https://github.com/dscripka/openWakeWord/releases/download/v0.5.1/hey_jarvis_v0.1.onnx lands in voice/openwakeword-v0.5.1

openWakeWord pre-trained ONNX model (the wheel downloads these at first run; offline needs them on disk)url · alexa_v0.1.onnx

source https://github.com/dscripka/openWakeWord/releases/download/v0.5.1/alexa_v0.1.onnx lands in voice/openwakeword-v0.5.1

openWakeWord pre-trained ONNX model (the wheel downloads these at first run; offline needs them on disk)url · hey_mycroft_v0.1.onnx

source https://github.com/dscripka/openWakeWord/releases/download/v0.5.1/hey_mycroft_v0.1.onnx lands in voice/openwakeword-v0.5.1

openWakeWord pre-trained ONNX model (the wheel downloads these at first run; offline needs them on disk)url · hey_rhasspy_v0.1.onnx

source https://github.com/dscripka/openWakeWord/releases/download/v0.5.1/hey_rhasspy_v0.1.onnx lands in voice/openwakeword-v0.5.1

Kokoro-82M v1.0 English ONNX, int8 (CPU) for the kokoro-onnx package; v1.1 is the zh modelurl · kokoro-v1.0.int8.onnx

source https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/kokoro-v1.0.int8.onnx lands in voice/kokoro-onnx-v1.0

Kokoro-82M v1.0 English ONNX, fp32url · kokoro-v1.0.onnx

source https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/kokoro-v1.0.onnx lands in voice/kokoro-onnx-v1.0

Kokoro v1.0 voice embeddings for kokoro-onnxurl · voices-v1.0.bin

source https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/voices-v1.0.bin lands in voice/kokoro-onnx-v1.0

IRON / MLIR-AIE wheel for XDNA2 NPU kernels (halo phases 1-3)findlinks · mlir_aie

source https://github.com/Xilinx/mlir-aie/releases/expanded_assets/latest-wheels-2

Peano: LLVM for AIE cores (wheel)findlinks · llvm-aie

source https://github.com/Xilinx/llvm-aie/releases/expanded_assets/nightly

NPU firmware the current amd/xdna-driver expects (17f0_11 = Strix Halo); halo ran driver 2.23 with dev fw 255.0.11.71whence · amdnpu-firmware

source https://raw.githubusercontent.com/amd/xdna-driver/main/tools/WHENCE

FastFlowLM 1.0.7 prebuilt flm runtime for Ubuntu 24.04 (30 MiB, sha256 4f11daad...dacd7; latest release, src tree is 61 commits ahead); depends on libxrt2 + libxrt-npu2 (lemonade-team PPA rows below) and noble ffmpeg/boost/fftw; weights are the FastFlowLM/*-NPU2 HF reposurl · fastflowlm_1.0.7_ubuntu24.04_amd64.deb

source https://github.com/ROCm/FastFlowLM/releases/download/v1.0.7/fastflowlm_1.0.7_ubuntu24.04_amd64.deb lands in npu/fastflowlm-v1.0.7

lemonade-team/stable PPA: XRT 2.25.0 base runtime (libxrt2), dependency of libxrt-npu2 and fastflowlm; version-pinned pool URL, ages out when the PPA moves onurl · libxrt2_2.25.0-4~noble1_amd64.deb

source https://ppa.launchpadcontent.net/lemonade-team/stable/ubuntu/pool/main/x/xrt/libxrt2_2.25.0-4~noble1_amd64.deb lands in npu/lemonade-ppa-noble

lemonade-team/stable PPA: XRT 2.25.0 NPU (amdxdna) shim, packaged alternative to building XRT from source; fastflowlm depends on iturl · libxrt-npu2_2.25.0-4~noble1_amd64.deb

source https://ppa.launchpadcontent.net/lemonade-team/stable/ubuntu/pool/main/x/xrt/libxrt-npu2_2.25.0-4~noble1_amd64.deb lands in npu/lemonade-ppa-noble

lemonade-team/stable PPA: amdxdna NPU kernel driver as DKMS (7.0-rc1 snapshot 2026-03-10); needs dkms from nobleurl · amdxdna-dkms_7.0.0-rc1+git20260310.6b13cb8f4-noble1_amd64.deb

source https://ppa.launchpadcontent.net/lemonade-team/stable/ubuntu/pool/main/a/amdxdna-dkms/amdxdna-dkms_7.0.0-rc1+git20260310.6b13cb8f4-noble1_amd64.deb lands in npu/lemonade-ppa-noble

Ryzen AI 1.8 Linux NPU driver package (86 MiB, public): XRT 2.25.37 base/base-dev/npu debs + xrt_plugin amdxdna deb, the XRT matched to the voe/onnxruntime-vitisai 1.8 wheels; the ryzen_ai-1.8.0.tgz installer itself is behind an AMD EULA formurl · RAI_1.8_Linux_NPU_XRT.zip

source https://download.amd.com/opendownload/RyzenAI/Driver/RAI_1.8_Linux_NPU_XRT.zip lands in npu/ryzenai-1.8-linux

Ryzen AI 1.8 Linux: ONNX Runtime 1.27 with the Vitis AI EP for the XDNA2 NPU (AMD public index, no login); needs voe==1.8.0findlinks · onnxruntime-vitisai

source https://pypi.amd.com/ryzenai_llm/1.8.0/linux/simple/onnxruntime-vitisai/

Ryzen AI 1.8 Linux: VOE (1 GB) = Vitis AI EP libs, vaip_config.json, xcompiler, dyn_bins (NPU binaries) for running ONNX models on the XDNA2 NPU; pins transformers 4.57.6, accelerate 1.12.0, datasets 4.5.0, evaluate 0.4.6findlinks · voe

source https://pypi.amd.com/ryzenai_llm/1.8.0/linux/simple/voe/

VoiceFixer restorer checkpoint (CC-BY-4.0, md5 1f479dabb9fa5724de5e20f2164cbc7a); the code wants it at ~/.cache/voicefixer/analysis_module/checkpoints/vf.ckpt. The weights were never on Hugging Face (old row haoheliu/voicefixer was a 404)url · vf.ckpt

source https://zenodo.org/records/5600188/files/vf.ckpt lands in voicefixer

VoiceFixer 44.1 kHz vocoder (CC-BY-4.0, md5 ac6de16561dd6042daf8bb3a806dc818); the code wants it at ~/.cache/voicefixer/synthesis_module/44100/model.ckpt-1490000_trimed.pturl · model.ckpt-1490000_trimed.pt

source https://zenodo.org/records/5600188/files/model.ckpt-1490000_trimed.pt lands in voicefixer

Intel oneAPI runtime pinned by torch 2.12.0+xpu (Arc Battlemage XPU lane; torch/torchvision/triton-xpu wheels already on the island)findlinks · intel-cmplr-lib-rt==2025.3.2

source https://download.pytorch.org/whl/xpu/intel-cmplr-lib-rt/

Intel oneAPI runtime pinned by torch 2.12.0+xpu (Arc Battlemage XPU lane; torch/torchvision/triton-xpu wheels already on the island)findlinks · intel-cmplr-lib-ur==2025.3.2

source https://download.pytorch.org/whl/xpu/intel-cmplr-lib-ur/

Intel oneAPI runtime pinned by torch 2.12.0+xpu (Arc Battlemage XPU lane; torch/torchvision/triton-xpu wheels already on the island)findlinks · intel-cmplr-lic-rt==2025.3.2

source https://download.pytorch.org/whl/xpu/intel-cmplr-lic-rt/

Intel oneAPI runtime pinned by torch 2.12.0+xpu (Arc Battlemage XPU lane; torch/torchvision/triton-xpu wheels already on the island)findlinks · intel-sycl-rt==2025.3.2

source https://download.pytorch.org/whl/xpu/intel-sycl-rt/

Intel oneAPI runtime pinned by torch 2.12.0+xpu (Arc Battlemage XPU lane; torch/torchvision/triton-xpu wheels already on the island)findlinks · dpcpp-cpp-rt==2025.3.2

source https://download.pytorch.org/whl/xpu/dpcpp-cpp-rt/

Intel oneAPI runtime pinned by torch 2.12.0+xpu (Arc Battlemage XPU lane; torch/torchvision/triton-xpu wheels already on the island)findlinks · intel-opencl-rt==2025.3.2

source https://download.pytorch.org/whl/xpu/intel-opencl-rt/

Intel oneAPI runtime pinned by torch 2.12.0+xpu (Arc Battlemage XPU lane; torch/torchvision/triton-xpu wheels already on the island)findlinks · intel-openmp==2025.3.2

source https://download.pytorch.org/whl/xpu/intel-openmp/

Intel oneAPI runtime pinned by torch 2.12.0+xpu (Arc Battlemage XPU lane; torch/torchvision/triton-xpu wheels already on the island)findlinks · mkl==2025.3.1

source https://download.pytorch.org/whl/xpu/mkl/

Intel oneAPI runtime pinned by torch 2.12.0+xpu (Arc Battlemage XPU lane; torch/torchvision/triton-xpu wheels already on the island)findlinks · onemkl-license==2025.3.1

source https://download.pytorch.org/whl/xpu/onemkl-license/

Intel oneAPI runtime pinned by torch 2.12.0+xpu (Arc Battlemage XPU lane; torch/torchvision/triton-xpu wheels already on the island)findlinks · onemkl-sycl-blas==2025.3.1

source https://download.pytorch.org/whl/xpu/onemkl-sycl-blas/

Intel oneAPI runtime pinned by torch 2.12.0+xpu (Arc Battlemage XPU lane; torch/torchvision/triton-xpu wheels already on the island)findlinks · onemkl-sycl-dft==2025.3.1

source https://download.pytorch.org/whl/xpu/onemkl-sycl-dft/

Intel oneAPI runtime pinned by torch 2.12.0+xpu (Arc Battlemage XPU lane; torch/torchvision/triton-xpu wheels already on the island)findlinks · onemkl-sycl-lapack==2025.3.1

source https://download.pytorch.org/whl/xpu/onemkl-sycl-lapack/

Intel oneAPI runtime pinned by torch 2.12.0+xpu (Arc Battlemage XPU lane; torch/torchvision/triton-xpu wheels already on the island)findlinks · onemkl-sycl-rng==2025.3.1

source https://download.pytorch.org/whl/xpu/onemkl-sycl-rng/

Intel oneAPI runtime pinned by torch 2.12.0+xpu (Arc Battlemage XPU lane; torch/torchvision/triton-xpu wheels already on the island)findlinks · onemkl-sycl-sparse==2025.3.1

source https://download.pytorch.org/whl/xpu/onemkl-sycl-sparse/

Intel oneAPI runtime pinned by torch 2.12.0+xpu (Arc Battlemage XPU lane; torch/torchvision/triton-xpu wheels already on the island)findlinks · tbb==2022.3.1

source https://download.pytorch.org/whl/xpu/tbb/

Intel oneAPI runtime pinned by torch 2.12.0+xpu (Arc Battlemage XPU lane; torch/torchvision/triton-xpu wheels already on the island)findlinks · tcmlib==1.4.1

source https://download.pytorch.org/whl/xpu/tcmlib/

Intel oneAPI runtime pinned by torch 2.12.0+xpu (Arc Battlemage XPU lane; torch/torchvision/triton-xpu wheels already on the island)findlinks · umf==1.0.3

source https://download.pytorch.org/whl/xpu/umf/

Intel oneAPI runtime pinned by torch 2.12.0+xpu (Arc Battlemage XPU lane; torch/torchvision/triton-xpu wheels already on the island)findlinks · intel-pti==0.16.0

source https://download.pytorch.org/whl/xpu/intel-pti/

Intel oneAPI runtime pinned by torch 2.12.0+xpu (Arc Battlemage XPU lane; torch/torchvision/triton-xpu wheels already on the island)findlinks · oneccl==2021.17.2

source https://download.pytorch.org/whl/xpu/oneccl/

Intel oneAPI runtime pinned by torch 2.12.0+xpu (Arc Battlemage XPU lane; torch/torchvision/triton-xpu wheels already on the island)findlinks · oneccl-devel==2021.17.2

source https://download.pytorch.org/whl/xpu/oneccl-devel/

Intel oneAPI runtime pinned by torch 2.12.0+xpu (Arc Battlemage XPU lane; torch/torchvision/triton-xpu wheels already on the island)findlinks · impi-rt==2021.17.2

source https://download.pytorch.org/whl/xpu/impi-rt/

Intel compute-runtime 26.35.39758.10 Level Zero GPU driver (xe KMD, Arc Battlemage production; sha256 c19a641b...0e88470). noble's 23.43 libze-intel-gpu1 is i915-only and finds no XPU on the RMA box's Arc B-series G31 (8086:e223)url · libze-intel-gpu1_26.35.39758.10-0_amd64.deb

source https://github.com/intel/compute-runtime/releases/download/26.35.39758.10/libze-intel-gpu1_26.35.39758.10-0_amd64.deb lands in gpu/intel-compute-runtime-26.35

Intel compute-runtime 26.35.39758.10 OpenCL ICD (sha256 61712caa...4ec4301), same release as libze-intel-gpu1url · intel-opencl-icd_26.35.39758.10-0_amd64.deb

source https://github.com/intel/compute-runtime/releases/download/26.35.39758.10/intel-opencl-icd_26.35.39758.10-0_amd64.deb lands in gpu/intel-compute-runtime-26.35

Intel gmmlib 22.10.0 as shipped with compute-runtime 26.35 (sha256 6031a63d...0e54d54); noble has 22.3.17url · libigdgmm12_22.10.0_amd64.deb

source https://github.com/intel/compute-runtime/releases/download/26.35.39758.10/libigdgmm12_22.10.0_amd64.deb lands in gpu/intel-compute-runtime-26.35

Intel Graphics Compiler 2.41.5 core (sha256 0a6e64a6...aaf4f628), required by compute-runtime 26.35url · intel-igc-core-2_2.41.5+22716_amd64.deb

source https://github.com/intel/intel-graphics-compiler/releases/download/v2.41.5/intel-igc-core-2_2.41.5+22716_amd64.deb lands in gpu/intel-compute-runtime-26.35

Intel Graphics Compiler 2.41.5 OpenCL front end (sha256 779e1b9e...1b87058), required by compute-runtime 26.35url · intel-igc-opencl-2_2.41.5+22716_amd64.deb

source https://github.com/intel/intel-graphics-compiler/releases/download/v2.41.5/intel-igc-opencl-2_2.41.5+22716_amd64.deb lands in gpu/intel-compute-runtime-26.35

oneAPI Level Zero loader 1.32.0 for Ubuntu 24.04, the loader compute-runtime 26.35 is validated with (noble ships 1.16.1); torch 2.12.0+xpu's UR adapter loads iturl · libze1_1.32.0+u24.04_amd64.deb

source https://github.com/oneapi-src/level-zero/releases/download/v1.32.0/libze1_1.32.0+u24.04_amd64.deb lands in gpu/intel-compute-runtime-26.35

HRX native bundle for hrx-rs 8f0e084 (libloomc/libamdf/libhrx_fabric, gfx1151, Ubuntu 26.04 build; sha256 57fa0d8509b7b495076feb7556edfb80e92829eb0b832788ce663e0758d60d68); offline: HRX_OFFLINE=1 hrx prepare <tarball>url · hrx-linux-x86_64-gfx1151.tar.gz

source https://github.com/zacharydenton/hrx-rs/releases/download/native-20260929-fbbf300312-fix1/hrx-linux-x86_64-gfx1151.tar.gz lands in hrx/native-20260929-fbbf300312-fix1

v0.7.8 Vulkan portable (llama.cpp 9e21f2e53 + Mesa 26.2.4, 40 MB): +22..46% prefill vs v0.7.6.1; run via top-level launchers (bin/ skips the bundled driver)url · strix-halo-llamacpp-vulkan-portable.tar.gz

source https://github.com/Nathanw1014/strix-halo-llamacpp/releases/download/v0.7.8/strix-halo-llamacpp-vulkan-portable.tar.gz lands in binaries/strix-halo-llamacpp/v0.7.8

v0.7.8 build manifest (source commit, Mesa, validation)url · MANIFEST.txt

source https://github.com/Nathanw1014/strix-halo-llamacpp/releases/download/v0.7.8/MANIFEST.txt lands in binaries/strix-halo-llamacpp/v0.7.8

Lemonade 2026.41.1 embeddable server tarball for Ubuntu x64 (7 MB): run lemonade-server on pod-control without the deb/rpm packagingurl · lemonade-embeddable-2026.41.1-ubuntu-x64.tar.gz

source https://github.com/lemonade-sdk/lemonade/releases/download/v2026.41.1/lemonade-embeddable-2026.41.1-ubuntu-x64.tar.gz lands in inference/lemonade-v2026.41.1

Lemonade 2026.41.1 server deb (debian13 build; check deps on noble before installing)url · lemonade-server_2026.41.1-debian13_amd64.deb

source https://github.com/lemonade-sdk/lemonade/releases/download/v2026.41.1/lemonade-server_2026.41.1-debian13_amd64.deb lands in inference/lemonade-v2026.41.1

Lemonade 2026.41.1 source tarballurl · lemonade-2026.41.1.tar.gz

source https://github.com/lemonade-sdk/lemonade/releases/download/v2026.41.1/lemonade-2026.41.1.tar.gz lands in inference/lemonade-v2026.41.1

lemonade-sdk prebuilt llama.cpp b1344 with ROCm 7 for gfx1151 (Strix Halo), Ubuntu x64, 480 MB: a ROCm reference build to A/B against /opt/llama (Vulkan) and the strix forkurl · llama-b1344-ubuntu-rocm-gfx1151-x64.zip

source https://github.com/lemonade-sdk/llamacpp-rocm/releases/download/b1344/llama-b1344-ubuntu-rocm-gfx1151-x64.zip lands in inference/llamacpp-rocm-b1344

Blender 5.2.2 Linux x64 release build (383 MB, sha256 84098912...06a168): the Mesh2Motion rig/animation .blend files are saved by Blender 5.0 and the mocap add-on wants 4.2+, noble ships 4.0.2; building blender/blender from source needs its multi-GB precompiled lib-linux_x64 tree, so the release tarball is the derivable-enough form (see docs/MESH2MOTION.md)binaries · blender

source https://download.blender.org/release/Blender5.2/blender-5.2.2-linux-x64.tar.xz

Rust 1.99.1 stable standalone installer (rustc+cargo+std, 211 MB, sha256 231fcae4...76cfe7): noble's rustc 1.75 is below Tauri 2's MSRV and fetch/cargo.sh needs cargo >= 1.85; lands in $ARK_SOFTWARE/binaries/ where fetch/cargo.sh looks for rust-*.tar.*binaries · .

source https://static.rust-lang.org/dist/rust-1.99.1-x86_64-unknown-linux-gnu.tar.xz

Node.js 24.22.0 LTS Linux x64 tarball (sha256 cc2f601c...e95008): the 'mirrored Node 24' docs/LESSONS.md assumes (noble ships nodejs 18); mesh2motion-app pins 24.11.0 in .nvmrc and Dockerfile, vite 8 / vitest 4 need >= 20binaries · node

source https://nodejs.org/dist/v24.22.0/node-v24.22.0-linux-x64.tar.xz

tauri-bundler downloads this at `tauri build` time for the appimage target (crates/tauri-bundler/src/bundle/linux/appimage/linuxdeploy.rs, commit hash 07333c6 hard-coded); offline: drop it in ~/.cache/tauri/ as linuxdeploy-07333c6-x86_64.AppImagebinaries · tauri-tools

source https://github.com/tauri-apps/binary-releases/releases/download/linuxdeploy-07333c6/linuxdeploy-x86_64.AppImage

tauri-bundler downloads this at build time for the appimage target; offline: ~/.cache/tauri/AppRun-x86_64binaries · tauri-tools

source https://github.com/tauri-apps/binary-releases/releases/download/apprun-old/AppRun-x86_64

optional: tauri-bundler tries this at build time and falls back to linuxdeploy's built-in plugin if the download fails; offline: ~/.cache/tauri/linuxdeploy-plugin-appimage.AppImagebinaries · tauri-tools

source https://github.com/linuxdeploy/linuxdeploy-plugin-appimage/releases/download/continuous/linuxdeploy-plugin-appimage-x86_64.AppImage

nfpm 2.48.0 Linux x86_64 (6.9 MB): the .deb/.rpm packager the ArtCraft Craft suite release pipeline uses (craft-common-actions install-nfpm); build deps list in verify/apt-artcraft.txtbinaries · nfpm

source https://github.com/goreleaser/nfpm/releases/download/v2.48.0/nfpm_2.48.0_Linux_x86_64.tar.gz

CC0 assets3 sources

polyhaven · hdrisCC0

selection all, 2k .hdr

polyhaven · texturesCC0

selection top 400 by downloads, 2k jpg maps (Diffuse nor_gl Rough AO Displacement arm Metal)

ambientcg · MaterialCC0

selection top 400 popular, 2K-JPG zip

Container images14

gitea/gitea:latest

note self-hosted git for the agents

qdrant/qdrant:latest

note vector db

chromadb/chroma:latest

note vector db

quay.io/jupyter/scipy-notebook:latest

note notebooks

pgvector/pgvector:pg17

note postgres + vectors

postgres:17

redis:7

nginx:stable

ollama/ollama:latest

note model runner

ghcr.io/open-webui/open-webui:main

note chat front

infiniflow/ragflow:nightly-slim

note RAG app

valkey/valkey:8

ghcr.io/peonist-ai/halogen-flash-server:0.16.0

note halogen: closed-binary engine for Qwen3.8-Flash-Next on Strix Halo gfx1151 only (ROCm bundled, ~0.9 GB); loads unsloth UD-IQ4_XS GGUF losslessly + its own 1.4 GiB draft head; EULA permits unmodified mirroring; wants the whole box (12 GB left)

node:24-alpine

note base image of mesh2motion-app/Dockerfile (pins node:24.11.0-alpine; latest 24 kept) so the documented `docker compose up` path works offline; generally useful for any Vite/Node front end

Extra PyPI packages154

Pulled beyond the repos' own requirements, into every wheel lane that has a build.

a2a-sdk

accelerate

ag2

agency-swarm

agentscope

agno

ai-edge-litert

ai-toolkit

anthropic

anticipation

anyjev

audio-separator

audioldm2

axolotl

camel-ai

capstone

chatdev

chromadb

cocotb

cocotb-bus

colpali-engine

composio-core

compressed-tensors

ctranslate2

datasets

datatrove

decider-models

deepagents

deepfilternet

deepspeed

demucs

distilabel

docling

dspy

einops

esptool

ethos-u-vela

evaluate

exllamav3

faiss-cpu

faster-whisper

fastmcp

FlagEmbedding

flash-linear-attention[rocm]

freenect

gfpgan

gptqmodel

graphrag

haystack-ai

heretic-llm

insightface

inspect-ai

jupyterlab

kev

kokoro

kokoro-onnx

lancedb

langchain

langchain-community

langchain-core

laya

liger-kernel

lighteval

lightrag-hku

litgpt

livekit-plugins-silero

llamafactory

llmcompressor

lm_eval

marimo

marker-pdf

matchering

matchering-cli

mcp-agent

mcp-use

mergekit

metagpt

miditoolkit

mido

mistral-common

mkdocs-material

mlflow

montreal-forced-aligner

moshi

moshi-mlx

ms-swift

music21

musubi-tuner

nemoguardrails

nerfstudio

ollaya

onnx2tf

onnxruntime

onnxsim

openrlhf

openwakeword

optimum

optimum-quanto

optuna

owl

peft

piper-tts

platformio

praisonai

pretty_midi

pyannote.audio

pyaudio

pycolmap

pyloudnorm

pymupdf

pyrealsense2

pyserial

pyuvm

qdrant-client

realesrgan

RealtimeSTT

RealtimeTTS

rembg

resemble-enhance

seed-vc

semantic-kernel

sentence-transformers

sherpa-onnx

silero-vad

sounddevice

sphn

strands-agents

swarms

swe-rex

symusic

tensorboard

tflite-runtime

tinytroupe

torchao

torchtune

trl

txtai

ultralytics

unsloth

unsloth[huggingface]

unsloth_zoo

unstructured

useful-moonshine-onnx

verl

voicefixer

vunit_hdl

wandb

webrtcvad

whisperx

wyoming

wyoming-faster-whisper

wyoming-openwakeword

wyoming-piper

wyoming-satellite

apt and interpreters3 notes

apt sources mirrored latest-only (see fetch/apt-mirror.py): noble main restricted universe multiverse (+updates/security)

extra pools: ROCm (repo.radeon.com), docker, kitware, graphics-drivers ppa, nvidia-container, cuda-repo-12-8-local

python-build-standalone: cpython 3.12/3.13/3.14 x86_64, x86_64_v3, aarch64 install_only_stripped + SHA256SUMS