| --- |
| license: other |
| license_name: minimax-h3-community-license-agreement |
| license_link: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE |
| base_model: MiniMaxAI/MiniMax-H3 |
| base_model_relation: quantized |
| library_name: mlx |
| tags: |
| - mlx-serve |
| - text-to-video |
| - image-text-to-video |
| - audio-video-generation |
| --- |
| |
| # MiniMax-H3 FL2VA — MLX-Serve 8-bit |
|
|
| 8-bit affine (group size 64) conversion of MiniMax-H3's FL2VA checkpoint for |
| [mlx-serve](https://github.com/ddalcu/mlx-serve), running natively on Apple |
| Silicon. Text-to-audio-video: the DiT denoises video and stereo audio jointly |
| in one packed sequence. |
|
|
| Self-contained: weights, both VAEs and the tokenizer in one directory. Upstream |
| splits these across `Comfy-Org/MiniMax-H3` (weights, no tokenizer) and |
| `MiniMaxAI/MiniMax-H3` (tokenizer). |
|
|
| Quantized: the DiT and text-encoder matmul weights. Kept dense: gathered |
| embedding tables, the checkpoint's fp32 islands (patch projections, output |
| heads, time embedder) and both VAEs. |
|
|
| Note that quantization here buys FOOTPRINT, not speed — the workload is |
| compute-bound at roughly 192,000 FLOPs per weight byte. |
|
|
| ## Modifications |
|
|
| These are MODIFIED files. The transformer and text encoder are quantized to |
| 8-bit; see MODIFICATIONS.md for the full list. The VAEs and tokenizer are |
| unmodified copies. |
|
|
| ## Turbo LoRA |
|
|
| `turbo_lora.safetensors` is included. It is a 4-step distillation adapter |
| trained by [larryvrh](https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora) |
| and licensed Apache-2.0, copied byte for byte from that repo's |
| `minimax_h3_turbo_4step_ema_ckpt850.safetensors` and renamed so mlx-serve finds it |
| inside the pack. Nothing about it was changed. |
|
|
| It samples in 4 to 8 steps instead of the usual 16 to 30 (larryvrh notes 4 is |
| already sharp on this checkpoint; earlier ones needed 6 to 8). Measured on an M4 Max |
| at 960x544, 39 frames, one prompt: 188s of sampling at 16 steps without it |
| against 95s at 4 steps with it, so roughly 2x on sampling and 1.8x on the whole |
| request once staging and VAE decode are counted. Per step it is about 20% |
| slower, because the adapter is summed at runtime rather than folded into the |
| weights and it turns off mlx-serve's step-caching recipe. The win is the step |
| count, not the step. |
|
|
| Treat it as a preview. This is the EMA weighting of checkpoint 850, which |
| larryvrh recommends as the cleanest of the round; the non-EMA file of the same |
| step is sharper still but over-sharpened, and their notes describe plastic skin |
| and over-sharp grain as the round's known failure modes. Their guidance if you |
| hit either: nudge LoRA strength down to 0.8-0.95 for grain, up to 1.05-1.2 for |
| blurry ghosting. |
|
|
| Request it with `"turbo": true`, or tick Turbo in the mlx-serve app's video |
| pane. It stacks with your own style LoRAs rather than replacing them. |
|
|
| ## License |
|
|
| Powered by MiniMax H3. Licensed under the MiniMax H3 Community License |
| Agreement -- see LICENSE and NOTICE, both included here. |
|
|
| **Territorial restriction.** The Agreement defines the Applicable Territory as |
| worldwide EXCLUDING the European Union, the United Kingdom, the Republic of |
| Korea and the United States of America, and Section V.4 prohibits use, |
| reproduction, modification, distribution and display outside it. Check whether |
| your jurisdiction permits you to use these files before downloading them. |
|
|
