MiniMax H3 open weights land: a 33B video model with native audio
The open reply to closed Seedance 2.5: #1 in video editing, 32kHz stereo baked in, and an 8-bit build that renders on a 12GB RTX 3060 — mind the license.
Copy markdown
What dropped: open weights, native audio
MiniMax put H3's weights on Hugging Face — a 33B model (~13B active) that generates 4–15s clips at up to 2K/24fps with 32kHz stereo audio synced to the picture. It's the self-hostable answer to ByteDance's closed Seedance 2.5, which shipped the same week.
It runs on a 12GB 3060
An 8-bit build renders 832×480 at 124 frames in under 10 minutes on an RTX 3060 (12GB) with 32GB RAM and an NVMe SSD — disk I/O sits on the inference path. Diffusers, SGLang, vLLM, and ComfyUI recipes ship in the repo.
#1 in video editing
On Artificial Analysis, H3 takes #1 in Video Editing (Elo 1130) and #2 in Text-to-Video (Elo 1242), just behind Gemini Omni Flash. Its Omni-Reference intake pins identity, motion, and sound from up to 9 images, 3 clips, and 3 audio files at once.
The license isn't MIT
H3 ships under the MiniMax Community License: free for non-commercial use and for companies under $20M revenue with prominent attribution; past that you negotiate. Read the license Q&A before you build a product on it.
Or skip the GPU: ~$0.13/sec
Don't want to self-host? The hosted API runs about $0.13/sec at 2K — a 15s clip ≈ $1.95 — via platform.minimax.io, fal.ai, and OpenRouter, roughly a third of Seedance 2.0's 1080p rate, with a 768p tier in beta.