| license: other | |
| license_name: minimax-h3-license | |
| license_link: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE | |
| base_model: MiniMaxAI/MiniMax-H3 | |
| tags: [svdquant, w4a4, nvfp4, video, text-to-video, quantized] | |
| # MiniMax-H3 · SVDQuant NVFP4 (rank 32, RTN) — reference format | |
| The RTX 50-series (sm_120) sibling of | |
| [the int4 release](https://huggingface.co/ModelsLab/MiniMax-H3-svdquant-int4_r32): | |
| e2m1 4-bit residual + bf16 rank-32 low-rank branch, quantized fresh from | |
| BF16 (int4 and fp4 grids do not nest, so transcoding is never used). | |
| **Status: reference (unpacked) format.** Runs through svdquant's | |
| reference backend today; the packed fp4 export for the Blackwell | |
| tensor-core kernels is in progress. RTN rounding (fp4-GPTQ pending). | |
| Same license inheritance and credits as the int4 release. | |