MiniMax Open-Sources H3: An Omni-Modal Video Model with Native 2K Stereo Audio
赛博禅心 · wechat · 2026-08-03
MiniMax has officially open-sourced MiniMax-H3, a new universal video model. It is an omni-modal generation system designed to understand complex inputs consisting of text, images, video, and audio, capable of generating videos up to 2K resolution and 15 seconds long with native stereo sound.
Core Architecture & Capabilities
- H3-Base: The core is a 33B parameter single-stream Transformer that jointly predicts video and audio latents. The text encoder utilizes the pre-trained weights of Qwen3-VL-32B.
- H3-Context-IR: A preprocessing system designed to deeply parse and distill complex multi-modal instructions into a structured representation to enhance generation quality.
- H3-Regenerate-2K: Instead of a traditional super-resolution module, it feeds the 768p output along with the original context back into the model to regenerate a 2K high-resolution video.
Deployment & Availability
The model supports text-to-video, first/last frame generation, and multi-modal reference generation (Ref2VA). Currently, the H3-Base weights are fully open and compatible with mainstream frameworks like SGLang, vLLM, diffusers, and ComfyUI for local deployment. The Context-IR and Regenerate-2K modules are provided via API to reproduce the official end-to-end workflow.
Related event: MiniMax open-sources 33B multimodal video model H3(35 posts)→
More from Models
- First Impressions: Gemini 4 Argon 'Incredible Writer' Beats Opus 5.5 and GPT 6 Astra — soumitrashukla9 · 2026-10-16
- Trying Grok bot surfaces an unpaid Cursor balance, blocking account merging — haydendevs · 2026-10-16
- TensorFold 1.0.12 runs 120B vision model on 1-2 DGX Sparks, adds Mac image input — HankYeomans · 2026-10-16
- Leaked screenshot shows unannounced Gemini 4 Argon gated to Ultra and paid Pro subscribers — il_-bro · 2026-10-16
- Google: Gemini 4 Argon lands in Antigravity, code submissions up 35%, rollbacks down ~33% — tekbog · 2026-10-16
- Google's "ungrounded beliefs" classifier injections derail Opus 4.1's reasoning chain — liminal_bardo · 2026-10-16