| --- |
| license: apache-2.0 |
| base_model: |
| - OpenMOSS-Team/MOSS-Image-SFT |
| - Qwen/Qwen3.5-9B |
| tags: |
| - webgpu |
| - int4 |
| - moss |
| - text-to-image |
| --- |
| |
| # MossInt4 |
|
|
| Small int4 versions of the Moss image model and its Qwen3.5 9B text encoder, made for running fully |
| in the browser with WebGPU. Everything is kept small to fit the browser's limits: WebAssembly can |
| address at most 4 GB, WebGPU buffers are capped at 2 GB, and most GPUs have 8-12 GB of VRAM. The |
| files are streamed with HTTP range requests straight into GPU buffers and cached in the browser |
| (OPFS), so they never have to fit into WebAssembly memory as a whole. |
|
|
| They are used by the [StrictComfy](https://github.com/strict-lang/StrictComfy) browser page, which |
| runs 100% locally on your device with WebGPU, written in [Strict](https://strict-lang.org). The |
| downloadable GitHub version (int8, CUDA) is up to 10x faster. |
|
|
| ## Files |
|
|
| | File | Size | Content | |
| |---|---|---| |
| | `moss_image_sft_4step_int4_convrot.safetensors` | 3.0 GB | Moss Image SFT with the 4-step LoRA merged, for 4 and 2 steps (fastest, fastest2steps) | |
| | `moss_image_sft_dmd_int4_convrot.safetensors` | 3.0 GB | Moss Image SFT with the DMD turbo LoRA merged, for 8 steps (best, fast, faster) | |
| | `qwen3.5_9b_int4_convrot.safetensors` | 4.5 GB | The language model of Qwen3.5 9B as the text encoder (no vision, MTP or output head) | |
|
|
| ## Format |
|
|
| ComfyUI's `convrot_w4a4` layout: every linear layer is rotated with a 256-wide Hadamard transform, |
| stored as signed 4-bit codes (two per byte, `I8`) with one `F32` scale per row, chosen by a |
| clip search per row (mean relative row error 11.4 %). Norms and small tensors stay `BF16`/`F32`, |
| the Qwen embedding stays int8. The LoRAs were merged into the bf16 weights before quantizing; the |
| Qwen file was requantized from the int8 convrot release. Computation runs with 8-bit activations |
| (WebGPU `dot4I8Packed`), so the int4 weights mainly save download size and memory. |
|
|
| Compared with the int8 originals the images keep the same composition with small differences in |
| fine detail; the text encoder's context has a cosine similarity of 0.96 to int8. |
|
|
| ## License |
|
|
| Apache 2.0, like everything they are made from. These files are modified versions (LoRAs merged, |
| quantized to int4) of: |
| - [OpenMOSS-Team/MOSS-Image-SFT](https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT) |
| - [akhaliq/MOSS-Image-SFT-4-steps](https://huggingface.co/akhaliq/MOSS-Image-SFT-4-steps) (4-step LoRA) |
| - [lodestones/MOSS-Image-SFT-8-steps-dmd](https://huggingface.co/lodestones/MOSS-Image-SFT-8-steps-dmd) (DMD LoRA) |
| - [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) |
