MiniMax drops SOTA open-weights H3 video model
MiniMax open-sourced H3's SOTA video-model weights; within 48 hours the community built a distillation LoRA cutting sampling from 20 to 4–8 steps, ran H3 on $280 gaming GPUs and fully-offline MacBooks, and shipped GGUF/ComfyUI ports. vLLM-Omni added Day 0 support with an OpenAI-compatible video endpoint.
How this story unfolded
13 days · 11 reports · 142 community posts · 153 of 154 shown
- Jul 31
- Aug 1
- Aug 2
- Aug 3
- Aug 4
- Aug 5
- Aug 6
- Aug 7
- Aug 8
- Aug 9
- Aug 10
- Aug 11
- Aug 12
- Aug 13
MiniMax by email
Get an email when MiniMax has news
No news that day, no email.
More stories today
- Amp adds fast mode to GLM-5.2 low mode, agent media caching
- Redwood Research chief scientist predicts AI progress over next few years
- Making an Actually Fun 3D Game with AI
- VisionDepth3D creates 3D from 2D videos with AI depth mapping
- xAI SDK v1.18.0 adds grok-4.6 and xhigh reasoning_effort