unsloth/MiniMax-H3-FP8

"MiniMax H3 is licensed under the MiniMax H3 Community License Agreement,
Copyright (c) 2026 MiniMax. All Rights Reserved."

The weights here derive from MiniMax-H3 (https://huggingface.co/MiniMaxAI/MiniMax-H3).
LICENSE is the text MiniMax ships at
https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE, verbatim.

Modified files and the changes made to them (MiniMax H3 Community License Agreement,
Section III.2). Both files start from minimax_h3_fl2va_pruned_bf16.safetensors as
published by Comfy-Org (https://huggingface.co/Comfy-Org/MiniMax-H3), which is the
H3-Base FL2VA diffusion transformer with the dense adaLN modulation projection replaced
by a rank-8 factorization of the time-embedding curve. That pruning is Comfy-Org's work
and not Unsloth's.

  * MiniMax-H3-INT8.pt: the 200 main-block 2-D matmul weights
    (blocks.N.{attn.qkv_proj, attn.out_proj, mlp.fc1, mlp.fc2}.weight) quantized to
    signed 8-bit integers with per-output-channel absmax scales, which changes the
    numerics. No block-Hadamard rotation is applied. Every other tensor is copied
    through bit-identical at its source precision.
  * MiniMax-H3-FP8.pt: the same 200 weights quantized to float8 e4m3 with
    per-output-channel absmax scales, which changes the numerics. Every other tensor is
    copied through bit-identical at its source precision.

Both are serialized as torchao pre-quantized state dicts rather than as safetensors.

Not an official MiniMax or Comfy-Org product, and not endorsed by either.