AnalysisVisual AIOctober 6, 2026

FastH3 V2 runs on a single consumer GPU, DGX Spark or Mac

Read original source →haoailab.com

The 138 GiB video-with-audio model now fits one machine via NVFP4 (Blackwell/DGX Spark), FP8 (RTX 4090) and INT6 (Apple Silicon) quantization, with activation scales calibrated on 1,000 prompts. FastH3 Trim drops 8 of 50 transformer blocks, is 4.2× smaller than base H3 and runs in as little as 8 GB of GPU memory.

1 source

More stories today

Open the live feed