minimax-h3-fp8

MiniMax-H3 — fp8 weights

fp8_e4m3fn copies of the MiniMax-H3 DiT and Qwen3-VL text encoder, in diffusers layout, produced with per-layer casting: fp8 storage, bf16 compute, with norms, embeddings and modulation layers deliberately left at full precision. fp8_restore/*.safetensors holds exactly those tensors, because the diffusers modular loader has no load-as-stored mode and must be asked for fp8 wholesale; fp8_cast.json records which components were cast and with which skip list.

This repo is an overlay, not a full model: it carries only what does not exist upstream. Combine it with MiniMaxAI/MiniMax-H3 — the bf16 DiT, the VAEs, the index, scheduler/tokenizer/processor — into one directory:

from huggingface_hub import snapshot_download

root = "models/h3-diffusers"
snapshot_download("MiniMaxAI/MiniMax-H3", local_dir=root, allow_patterns=[
    "modular_model_index.json", "*.json", "transformer/*", "vae/*", "audio_vae/*",
    "scheduler/*", "audio_scheduler/*", "processor/*", "tokenizer/*",
], ignore_patterns=["text_encoder/*", "transformer_ref/*", "FL2VA/*", "Ref2VA/*", "docs/*"])
snapshot_download("SabiQG/minimax-h3-fp8", local_dir=root)

Weights derive from MiniMaxAI/MiniMax-H3 and remain under its license — read it before using them.

This FP8 build was created to help power some of the generative video pipelines used by Aificient Studio.