Anima-3.8B-q8-web-xpu-ops

Anima-3.8B, converted to int8 for web-xpu-ops

Read this section before the rest.

These are not my weights, and they are not MIT. They are a converted copy of somebody else's model, redistributed under a licence that permits it and imposes conditions. Non-commercial and non-production use only.

The software that produced them — web-xpu-ops, a WebGPU kernel library — is MIT. That MIT licence covers the code and covers nothing in this repository. Running MIT code against these files does not make your use of them non-commercial-exempt, and finding the two next to each other does not merge their terms.

Where these came from

nvidia/Cosmos-Predict2-2B-Text2Image     NVIDIA Open Model License
  └── circlestone-labs/Anima             CircleStone Labs Non-Commercial License v1.2
        └── lylogummy/Anima-3.8B         no licence declared; defers to its upstream
              └── this repository        int8 conversion of the above

Two licences apply at once and the more restrictive one governs. NVIDIA's permits commercial use; CircleStone's does not. The answer is non-commercial.

Full texts: LICENSE.md (CircleStone), LICENSE.nvidia.txt (NVIDIA), and NOTICE, which carries the attribution the licences require verbatim.

Any right you have to use these weights is granted to you directly by CircleStone Labs LLC under that licence — not by me.

Built on NVIDIA Cosmos.

This is not an official product of CircleStone Labs LLC, lylogummy, or NVIDIA, and none of them has endorsed, approved or validated it.

What was changed

The diffusion transformer only. Section 3(d)(i) of the CircleStone licence requires this to be stated, so it is stated precisely:

• weights quantized from f32 to per-row absmax int8, range [-127, 127], four codes packed per 32-bit word, least-significant byte first;
• one f32 scale per output row, applied after the dot product;
• self-attention query and key projections permuted to match this runtime's rotary-embedding layout;
• tensors that quantization would damage — the patch embedding and the final projection — left in f32, in dit.f32.bin.

Nothing was retrained or fine-tuned. The encoder and the VAE are byte-for-byte copies of the upstream files.

Quantization is lossy and the cost is measured, not assumed. Against the unquantized model over all 52 blocks the output moves by 4.018e-2 relative RMS, measured independently in PyTorch. This runtime reproduces the quantized model to 1.18e-5, which is the number that says the port is faithful; the 4.018e-2 is what int8 costs and is a property of the format, not of the port.

What this format is

A layout that a browser can Range-request and a WebGPU kernel can read without a dequantization pass: dit.manifest.json names every tensor with its shape and byte offset, dit.q8.bin holds the packed codes, dit.q8scales.bin the per-row scales, dit.f32.bin the tensors left dense. It is produced by examples/anima/tools/convert_dit.py and read by examples/anima and examples/anima-web.

It is not a general interchange format and nothing else reads it. If you want the model for any other purpose, take it from lylogummy/Anima-3.8B instead — you will get the original f32 weights rather than a lossy conversion.

Using it

git clone https://github.com/m96-chan/web-xpu-ops
cd web-xpu-ops && npm install
npx tsx examples/anima/src/generate.ts \
  --dit <this repo> \
  --encoder <this repo>/qwen_3_06b_base.safetensors \
  --vae <this repo>/qwen_image_vae.safetensors \
  --prompt "..." --out out.png

Needs a GPU with WebGPU. Measured on an RTX 5090: 225 s for 40 steps at 832x1216 in Chrome, 2.18 s per forward in Node.

Provenance of this description

Every licence claim above is quoted from or cited to a document that is linked. Nothing is a legal opinion, from me or from anyone. If you intend to build a product on this, read LICENSE.md yourself.