開放模型與影音生成
MiniMax H3 Releases 33B Base Weights with Native Synchronized Video and Stereo Audio Generation
MiniMax H3 uses a single Transformer to jointly predict video and audio latents, supporting generation from text, first and last frames, and multimodal references. The open release can generate 768p audiovisual content locally, but the official 2K workflow still relies on two proprietary cloud modules.
MiniMax has released two sets of BF16 base weights for [H3](https://huggingface.co/MiniMaxAI/MiniMax-H3): FL2VA accepts text with optional first and last frames, while Ref2VA can simultaneously accept up to nine images, three video clips, and three audio clips. Both can generate videos lasting 4 to 15 seconds at 24 FPS with 32 kHz stereo audio. The base output defaults to 768 pixels on the shorter side.
At its core, the H3-Omni-Transformer is a 33B-parameter dense, single-sequence model, with roughly 13B parameters in AdaLN branches that can be precomputed and cached. Text and visual semantics are extracted from layer 50 of Qwen3-VL-32B. Images and audio are then compressed by a temporally causal VisualVAE and an AudioVAE, respectively, after which the Transformer predicts both types of latent synchronously. Visual latents use spatial and temporal compression ratios of 16× and 4×, respectively, followed by `1×2×2` patchification; audio is compressed into 40 latent tokens per second.
Deployment is far from lightweight. [SGLang’s native pipeline](https://docs.sglang.io/cookbook/diffusion/MiniMax/MiniMax-H3) recommends four H200 GPUs, or four 80GB H100 GPUs configured with TP2 and Ulysses2. The model card also lists vLLM, Diffusers, and ComfyUI. The initial weights support only full attention; the sparse attention implementation used during training has yet to be released.
Engineering teams should pay particular attention to the scope of what is “open.” The multi-model orchestration service that converts free-form inputs into structured Context-IR, as well as the module that regenerates 768p results together with the original conditioning inputs at 2K resolution, are available only through APIs. In other words, H3-Base can be reproduced locally, but the complete official 2K system cannot. The model uses a custom community license, so commercial deployment requires reviewing its geographic and use-case restrictions. The next points to watch are the sparse attention implementation, end-to-end VRAM usage and latency figures, and independent blind tests of audiovisual synchronization, character consistency, and text rendering.