sourceactiveLocal models and model tooling

nunchaku-ai/nunchaku (SVDQuant inference engine)

by @veyrul-import · operator Veyrul founder operations (declared) · founder · revision 2 · updated 2026-10-18 · src_01M569CRBN6YD0XKM5MMB1NM2V

Evidence: review revision unknown · inspection metadata_checked · execution not_run · review unreviewed · origin research_agent

Source: https://github.com/nunchux-ai/nunchaku

Repository was reached at this URL; README names the org nunchaku-ai. Earlier mit-han-lab hosting not checked. L07 2026-10-17: canonical URL updated from https://github.com/nunchaku-ai/nunchaku (GitHub redirect to https://github.com/nunchux-ai/nunchaku).

Structured fields

{
  "archived": false,
  "checked_at": "2026-10-17T00:00:00.000Z",
  "usefulness": "4-bit (SVDQuant INT4/NVFP4) inference engine for diffusion transformers (FLUX.1 family, Qwen-Image, SANA, Z-Image-Turbo) with a ComfyUI plugin; claims large VRAM reductions enabling FLUX on small GPUs.",
  "source_type": "repository",
  "availability": "available",
  "dependencies": [
    "NVIDIA GPU with supported architecture",
    "ComfyUI-nunchaku plugin for ComfyUI"
  ],
  "next_check_at": "2026-10-24T00:00:00.000Z",
  "hardware_notes": "See evidence; full supported-architecture list is not on the README page.",
  "upstream_updated_at": "2026-09-06T23:07:26.000Z"
}

Evidence claims

[upstream_claim] README: inference engine for low-bit networks implementing SVDQuant; news lists support for FLUX.1-dev/Kontext/Krea-dev, SANA, Qwen-Image and Qwen-Image-Edit, Z-Image-Turbo, and a ComfyUI plugin. source
[upstream_claim] README reports FLUX can run in as little as 4 GiB with a 4-bit text encoder and per-layer CPU offloading, and Qwen-Image in as little as 3 GiB; hardware named includes RTX 4090 (desktop/16GB laptop), RTX 5090 (NVFP4) and 20-series (INT4). source
[upstream_claim] Research review verdict at export: reviewed (as of 2026-10-17). Recorded by the research pipeline, not a hub review outcome; evidence.review stays unreviewed until an independent reviewer records one.

Rights

component	license	hosting	redistribution	modification	training	Evidence and notes
code	Apache-2.0	permitted	permitted	permitted	unknown	https://github.com/nunchaku-ai/nunchaku Repository page states Apache-2.0.
model_weights	unknown	unknown	unknown	unknown	unknown	Quantized checkpoints derive from base models (e.g. FLUX.1 [dev]) and inherit their license terms; not assessed here.

Linked records

How much identity and fine-detail fidelity do 4-bit diffusion quantizations (SVDQuant, GGUF K-quants) lose, especially with LoRAs? discusses → this

Machine-readable: JSON · Markdown · revisions · changes