多模態模型
MiniMax Opens H3 Base Weights: 33B Single-Stream Model Generates Stereo Video up to 15 Seconds
MiniMax H3 packages text, images, video, and audio into a single sequence and jointly predicts visual and audio latents, producing up to 768p video at 24 FPS with 32 kHz stereo audio. The official workflow claims regeneration at up to 2K, but prompt orchestration and the high-resolution module remain available only through the hosted API.
MiniMax has released two BF16 base checkpoints for H3, supporting text- or start/end-frame-conditioned generation and multimodal reference generation, respectively. The model can accept up to nine images, three video clips, and three audio clips, generating 4- to 15-second videos at 24 FPS with 32 kHz stereo audio. The standard output from the local weights has a 768-pixel short side.
H3-Base first uses dedicated encoders and VAEs to compress each modality into a unified sequence. A dense, 33B-parameter H3-Omni-Transformer then jointly predicts audiovisual latents. For text and visual understanding, it uses hidden states from layer 50 of Qwen3-VL-32B. The video VAE applies 16× spatial compression and 4× temporal compression, while the audio VAE converts each channel into 40 latent tokens per second. The Transformer itself does not create modality-specific branches in its attention or FFN layers; most modality-specific differences are confined to the input/output layers and AdaLN.
Notably, what MiniMax calls the complete H3 system is not equivalent to the weights currently available for download. H3-Context-IR, which parses free-form source material and generates structured prompts, and H3-Regenerate-2K, which regenerates 768p results at 2K using the original conditions, still depend on the MiniMax API. The sparse attention used during training has also not yet been released, so the initial inference implementation supports only full attention. As a result, the official SGLang example requires four GPUs for sequence-parallel deployment.
For engineering teams, this release provides a rare native model for synchronized audio-video generation, along with reproducible 768p examples. However, results from a locally deployed H3-Base should not be treated as representative of the official 2K product quality. Key developments to watch include the release of the sparse-attention implementation, whether the 2K regeneration module becomes genuinely open, and whether the community can reproduce audiovisual synchronization, text rendering detail, and long-shot consistency under identical prompt and hardware conditions.