MiniMax's H3 Beats Closed Labs in Video Editing With Open Weights

Video generation was the last major modality where closed labs still held a comfortable moat. Weights stayed private, quality stayed behind API keys, and frontier-quality clips came rented by the second. MiniMax H3 closes that gap. MiniMax announced the model on July 31 as an API-only product, then published the weights to Hugging Face on August 3 under the MiniMax H3 Community License. The community had it running on $280 gaming GPUs within a day.

The pitch is unusual for a video model. One unified context reads text, images, video, and audio together and generates coherent audiovisual results from any mix of them, covering text-to-video, first-and-last-frame, reference-to-video, and precise video editing in one checkpoint. The audio is generated jointly, not bolted on in a second stage.

One model, four modalities, one pass

The core spec sheet is where H3 stops looking incremental. It is a 33.1B dense, single-stream Omni Transformer, of which roughly 13B sits in AdaLN branches that can be precomputed and cached for inference-only deployment. The text encoder is Qwen3-VL-32B, position encoding is 3D multimodal RoPE across (t, h, w), and the model produces 4 to 15 second clips at 24 fps with 32 kHz native stereo audio generated in the same pass as the picture.

The reference budget per generation is generous by open-model standards: up to 9 reference images, 3 reference video clips, and 3 reference audio clips per generation. One shot can inherit a face from an image, a camera move from a video, and a voice from an audio clip, all steered by a single sentence prompt.

The four load-bearing tricks

MiniMax attributes the jump to four engineered pieces. Two of them matter for anyone thinking about inference cost: