MiniMax-H3 FP8 (TP4 / TP1 prepared weights)
Modified files. This repository contains Model Derivatives of MiniMaxAI/MiniMax-H3: the original weights were converted and quantised. They are NOT the original files.
MiniMax H3 is licensed under the MiniMax H3 Community License Agreement (see LICENSE and NOTICE), and so are these
weights. Read it before use: among other things it limits use, reproduction and distribution to its "Applicable
Territory" (worldwide excluding the European Union, the United Kingdom, the Republic of Korea and the United States),
requires machine-generated content to be disclosed, and has an Acceptable Use Policy (Exhibit A).
Contents
| folder | what | size |
|---|---|---|
base-fp8/ |
base text-to-video/audio DiT, FP8 E4M3 weights (codes capped at 224) + per-output-channel f32 scales, tensor-parallel 4 shards (rank0..3), bf16 token refiner, full-form AdaLN in host/ |
44 GB |
base-fp8-tp1/ |
the same DiT for one GPU (block scales [K/128, N]) | 20 GB |
ref2va-fp8/ |
Ref2VA (reference-to-video) DiT, same format as base-fp8/ |
44 GB |
text-encoder-fp8/ |
text encoder (Qwen3-VL based, 50 layers), FP8, TP4 | 25 GB |
vision/ |
vision tower for reference images, extracted from the Ref2VA text-encoder shard (bf16) | 1.2 GB |
vae-bf16/ |
video VAE decoder, bf16 | 4.6 GB |
Not included (use the original repo): audio VAE, Ref2VA video-VAE encoder, tokenizer / processor / scheduler configs.
Format
These are engine-specific prepared layouts (per-rank tensor files + meta.json), produced by the Minimax-neuron
prepare_weights tools (fp8 mode: per-channel minmax E4M3, max code 224; interleaved per-head qkv layout,
fc1 = [gate; value]). They load in that engine's Vulkan backend (KHR / NV cooperative matrix, TP over several GPUs)
and Neuron backend; they are not a diffusers / safetensors drop-in.
Measured with the Vulkan backend on 4 x RTX 5090 (TP4): 5.2 s 1344x768 text-to-video in ~2.5 min at 20 steps + FSampler; per-tensor quantisation error 2.6e-2; one real-width DiT block vs an fp32 reference: rel. L2 ~1e-2.
Generated videos must be disclosed as machine-generated (license, Exhibit A item 12).
Model tree for cgg507/MiniMax-H3-FP8-TP4
Base model
MiniMaxAI/MiniMax-H3