| --- |
| license: mit |
| library_name: mlx |
| pipeline_tag: text-to-video |
| tags: |
| - mlx |
| - apple-silicon |
| - video-generation |
| - text-to-video |
| - image-to-video |
| - video-continuation |
| - longcat |
| - flow-matching |
| - block-sparse-attention |
| - quantized |
| - 8-bit |
| base_model: |
| - mlx-community/LongCat-Video-bf16 |
| language: |
| - en |
| - zh |
| --- |
| |
| Part of the [LongCat-Video — MLX](https://huggingface.co/collections/mlx-community/longcat-video-mlx-6a216a3576c098e83c1cc167) collection. |
|
|
|
|
| # LongCat-Video-q8 (MLX) |
|
|
| 8-bit quantized variant of [mlx-community/LongCat-Video-bf16](https://huggingface.co/mlx-community/LongCat-Video-bf16). |
| Same model, same six task variants (T2V / I2V / Continuation / Refinement / Long-Video / Interactive), |
| same `cfg_step_lora` + `refinement_lora` files — just with the DiT Linears |
| quantized to 8-bit via `mlx.nn.quantize`. |
|
|
| The 8-bit variant trades a small disk-savings improvement (vs 4-bit) for |
| **near-bf16 quality**. If you have the RAM headroom for 30 GB but not 42 GB, |
| q8 is the right pick. |
|
|
| ## TL;DR |
|
|
| | | | |
| |---|---| |
| | **DiT** | 8-bit quantized (`group_size=64`, skip `final_layer.linear` + embedders + AdaLN) | |
| | **DiT size** | ~15 GB (4 shards; 1.7× smaller than bf16's 26 GB) | |
| | **VAE / umT5 / LoRAs** | bf16 (unchanged from bf16-variant) | |
| | **Total disk** | ~31 GB (vs 42 GB bf16) | |
| | **Min unified memory** | ~48 GB recommended for 480p | |
| | **Inference** | 50-step baseline OR 8-step with `cfg_step_lora` (fast) | |
| | **License** | MIT | |
|
|
| ## Quantization details |
|
|
| Same skip pattern as q4 — see the q4 card for full notes on why each |
| pattern is excluded (L11 + L42 in the |
| [skill-lessons](https://github.com/xocialize/longcat-video-mlx/blob/main/docs/development/skill-lessons.md)). |
|
|
| The only difference vs q4 is `bits=8` in the `quantization` config block. |
|
|
| ## Quick start |
|
|
| ```bash |
| # 1. Pull weights (~31 GB) |
| hf download mlx-community/LongCat-Video-q8 --local-dir ./weights |
| |
| # 2. Set up inference |
| git clone https://github.com/xocialize/longcat-video-mlx |
| cd longcat-video-mlx |
| python3.12 -m venv .venv |
| .venv/bin/pip install -e ".[parity]" |
| |
| # 3. Run text-to-video — pass --variant q8 |
| .venv/bin/python scripts/run_t2v.py \ |
| --weights ./weights/.. \ |
| --variant q8 \ |
| --prompt "A cat surfing on a wave at sunset, cinematic, 8k" \ |
| --num-frames 93 \ |
| --out output_t2v.mp4 |
| ``` |
|
|
| ## Choosing between bf16, q4, q8 |
|
|
| | Variant | Disk | Min RAM | Quality | Pick when | |
| |---|---|---|---|---| |
| | **bf16** | 42 GB | 64 GB | reference | Best output, you have the RAM headroom | |
| | **q4** | 25 GB | 32 GB | minor degradation | RAM is tight (32 GB Mac) | |
| | **q8** | 30 GB | 48 GB | very close to bf16 | Best balance — small savings, near-bf16 quality | |
|
|
| ## License |
|
|
| MIT — matches the upstream |
| [LongCat-Video](https://github.com/meituan-longcat/LongCat-Video) license. |
|
|
