| MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved. |
|
|
| FastVideo-FastH3-8-Step-V2 is a modified MiniMax H3 model from the FastVideo team. |
| The transformer |
| at training step 1300, not the original MiniMax H3 transformer weights. |
| The accompanying transformer configuration and index describe that student. |
| The scheduler configurations and fastvideo_inference.json describe its trained |
| eight-forward schedule, video/audio shifts 10/3, and 80% tile-64 VSA policy. |
| The modular model index, model card, and release/provenance documentation were |
| prepared by FastVideo for this derivative. |
|
|
| The bundled text encoder, tokenizer, processor, video VAE and audio VAE weights |
| are unmodified MiniMax H3 components. The Qwen3-VL encoder is licensed under |
| Apache License 2.0; a copy is included as LICENSE-Qwen3-VL. |
| Source: https://github.com/QwenLM/Qwen3-VL |
|
|
| The model license is provided in LICENSE. These notices do not replace or |
| modify its terms. |
|
|
| --- |
|
|
| fasth3-8-step-v2-fp8.rad is a modified copy of FastVideo-FastH3-8-Step-V2 |
| (FastVideo/FastVideo-FastH3-8-Step-V2), converted for the radiance inference |
| engine: the linears of the text encoder, the token refiner and the DiT and the |
| modulation projections are re-encoded as FP8 E4M3 with a bf16 scale per |
| 128 x 128 block, the video decoder's linears are cast from f32 to bf16, the |
| audio decoder's weight norms are folded, the text encoder's last 14 layers (not |
| read by the model) are left out, and the result is repacked into one .rad |
| container. minimax-h3-fp8.recipe lists the rules applied. |
|
|
| The latent upscaler is Minimax_h3_latent_Upscaler v1 by LBH-123-AI |
| (https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler), licensed under |
| the Apache License 2.0; a copy is included as LICENSE-latent-upscaler. Its |
| weights are included as published (bf16), its embedding MLP cast to f32. |
| |
