MiniMax H3 just got up to 14× FASTER. 🤯
And no, this isn't just another quant.
FastH3 combines:
⚡ Just 4 transformer forwards
🧠 VSA-H3 — 90% sparse attention
🎬 Video + synchronized audio
🔥 DMD2 distillation
Kijai has already brought the experimental version to ComfyUI, so I tested it locally on my RTX 5090.
30 steps → 4 steps.
The speed difference is ridiculous.
But what happens to the quality? 👀
Full review + benchmarks 👇