feverdoingwork/FastH3-Omni-8-Step-V1-W4A8

FastH3-Omni-8-Step-V1 - W4A8 (mixed) transformer

Modified-file notice: FastH3-Omni-8-Step-V1-pruned_rank8_w4a8_mixed.safetensors is a modified (quantized) derivative of FastVideo/FastVideo-FastH3-Omni-8-Step-V1 (MiniMax H3). It was quantized from the official bf16 file FastH3-Omni-8-Step-V1-pruned_rank8_bf16.safetensors (from DeepBeepMeep/MiniMax-H3, declared sha256 00aaa2d6e824d62f844409c3f37d8171dae0821fa2501350ed63c68e4564cf25).

Status: UNVERIFIED. Tensor-level and engine-handler checks pass (below). No end-to-end FastH3 Omni generation has been run with this file yet.

What it is

The transformer only, 14.64 GB (bf16: 44.25 GB). It keeps the 8-step schedule: the FastH3 sampling schedule is read from the file's fasth3_inference metadata, which is preserved unchanged. Quantizing does not change the step count.

layers	count	format
blocks.N.attn.qkv_proj, attn.out_proj, mlp.fc1, mlp.fc2	200	asym_w4a8_int8: 4-bit codebook weights, group size 16, ConvRot 256, FP8 E4M3 group scales
blocks.N.attn.gate_proj (VSA gate)	50	INT8 ConvRot (as in the INT8 release)
norms, adaLN, PDD heads, patch projections, token refiner	rest	copied unchanged (qkv rows regrouped to the quantized "grouped" layout)

Decode: int8 = clamp(rint(codebook[idx] * s_rel), -127, 127), W_rot = int8 * s_channel, W = inverse ConvRot (Hadamard 256). The 16-level codebook is the one used by existing W4A8 H3 checkpoints. Per 16-column group, s_rel is the FP8 value that minimises the squared error of that exact decode. See quantize_w4a8_h3.py and PROVENANCE.json.

Checks run

• Relative RMSE vs the bf16 source (rotated domain): mean 0.0719 (min 0.0716, max 0.0721) over the 200 W4A8 layers; cosine 0.9974.
• gate_proj INT8 layers are identical to the INT8 release; BF16 token_refiner qkv is bit-exact with it.
• The engine's asym_w4a8_int8 handler claims exactly the 200 W4A8 layers and int8_convrot the 50 gate layers; a Triton W4A8 forward on 3 sample layers matched bf16 within about 7.3% relative RMSE.

Use

This is a transformer checkpoint for engines with an asym_w4a8_int8 loader (WanGP / Zethus). Keep w4a8 in the file name: the engine labels the format from it. Use it with the standard FastH3 Omni companions (text encoder, video VAE, audio VAE, latent upscaler).

License

MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright (c) 2026 MiniMax. All Rights Reserved. See LICENSE and NOTICE. The license grants rights only within its "Applicable Territory" (worldwide excluding the European Union, the United Kingdom, the Republic of Korea and the United States) and binds recipients to its Acceptable Use Policy; read it before using or sharing these weights.