MiniMax releases open weights for its H3 video model
MiniMax published downloadable weights for H3, its omni-modal 2K video model, making a top-ranked video generator locally runnable for the first time.
MiniMax released the model weights for H3, its omni-modal video generation system, at midnight Beijing time on August 3 — 16:00 UTC on August 2 — following through on an open-sourcing pledge the Shanghai company made when it launched the model on July 31. The drop landed first on the domestic ModelScope hub, with Hugging Face mirrors following, and ComfyUI simultaneously published a local-inference tutorial covering the newly public checkpoints.
H3 is unusual among video models in treating text, images, video and audio as a single input context rather than as separate conditioning channels. It generates clips of roughly four to fifteen seconds at up to 2K resolution and 24 frames per second, and produces native stereo audio inside the same generation pass instead of dubbing sound on afterward. On Artificial Analysis leaderboards the model ranks first in video editing, second in text-to-video and third in image-to-video — placing it among the strongest video generators of any licensing type.
The published artifacts give an unusually clear view of the architecture. ComfyUI's local workflow pulls two pruned int8 diffusion checkpoints — one for text-to-video and image-to-video, one for reference-driven generation — plus separate video and audio VAEs and, notably, a text encoder derived from Alibaba's Qwen3-VL 32B vision-language model. Weights ship under the MiniMax Community License, which permits free non-commercial use and free commercial use by organisations below roughly $20 million in annual revenue, with attribution. MiniMax continues to sell hosted access through its own API and the Hailuo consumer app at prices it describes as under a third of comparable services per second of 2K output.
Why it matters
Until now the frontier of video generation has been almost entirely closed: OpenAI's Sora, Google's Veo and ByteDance's Seedance line are API-only. H3 breaks that pattern by putting a leaderboard-topping video model on local hardware, where studios can fine-tune it, inspect it, and run it without sending footage to a vendor. It also sharpens a strategic split inside China's video-model race — ByteDance widening closed API access to Seedance 2.5 the same week MiniMax gave its weights away — and hands the local-inference ecosystem, from ComfyUI workflows to quantisation projects, a genuinely competitive base model to build on.