MiniMax open-sources H3: 2K video with native audio in one pass
First open-weights model to render video and stereo audio in one pass, day-0 in ComfyUI. Also: Qwen3.8-Max's API and the open-frontier safety gap.
Copy markdown
Video and sound, generated together
H3 is a 33.1B-param omni transformer that renders up to 15s of 768p video (with optional 2K upscale) and 32 kHz stereo audio in a single forward pass — no bolt-on audio model. Weights are on Hugging Face under MiniMax's Community License, spanning 11 languages.
Day-0 in ComfyUI, and it fits your GPU
ComfyUI shipped native support at launch. Pruned INT8 checkpoints are ~40% smaller and an NVFP4 text encoder is meant to run on 'any GPU'; SGLang, vLLM and diffusers work too. This weekend: local lip-synced avatars or product clips with built-in SFX, no separate TTS or foley pass.
The hosted API undercuts Kling and Seedance ~3x
Don't want to self-host? H3's API is $0.13/s ($7.80/min) at 2K — versus roughly $20-22/min for Kling 3.0 and Seedance 2.0 at only 1080p. One gotcha: reference-video seconds get added to your billed output duration.
Elsewhere: Qwen3.8-Max's API opens
Alibaba turned on API access to Qwen3.8-Max, a ~2.4T-param, 1M-context flagship aimed at long-horizon agents. No open weights yet, so it's hosted-only for now.
Elsewhere: open weights near the frontier, safety trails
A SaferAI review (via TechCrunch) finds Z.ai's GLM-5.2 approaching frontier coding — beating GPT-5.5 on several long-horizon benchmarks at ~1/6th the cost — while shipping without comparable safety mitigations. Worth weighing if you're wiring open weights into autonomous agents.