FastH3 V2 native FastVideo Blackwell bundle

Unmodified native-format weights, configs and tokenizers for h3-runpod on one 96 GB RTX PRO 6000 Blackwell. Includes the NVFP4 attention/MLP/gate transformer, 50-layer NVFP4 text encoder, LynnReal 26-layer lightweight video VAE with INT8 ConvRot overlay, H3 audio VAE, and the original eight-forward DMD schedule. This bundle is for native FastVideo, not ComfyUI.

Source: FastVideo/FastVideo-FastH3-8-Step-V2-NVFP4-Consumer. Source revision: 9d0d0d3ae69365aded83fab5562f00f9f06c12aa. License copied from MiniMaxAI/MiniMax-H3 at 42ed227ee7df40d41602854ae760620d6eb651fe. See LICENSE and NOTICE for terms and attribution. No weights were modified. The upstream README is preserved in UPSTREAM_README.md.

Set the RunPod endpoint Model field to cokeadrink/fasth3-v2-native-blackwell. The worker discovers the mounted cache; no HF_MODEL_ID or HF_REVISION variables are needed.

Downloads last month: 31

Safetensors

Model size

14B params

Tensor type

BF16

Model tree for cokeadrink/fasth3-v2-native-blackwell

Base model

MiniMaxAI/MiniMax-H3

Finetuned

FastVideo/FastVideo-FastH3-8-Step-V2

Finetuned

(8)

this model