MiniMax Releases Open-Source H3: An Omni-Modal Model with Native Audio-Video Generation
aigclink · x · 2026-08-03
MiniMax has officially released and open-sourced the weights for H3, a universal omni-modal generation system. It can uniformly understand text, images, video, and audio, directly generating video content with native stereo sound.
Core Capabilities & Advantages:
- Unified Multimodality: Supports up to 15s of video generation at 2K resolution with native dual-channel audio. Features image-to-video, video-to-video (V2V) motion transfer, and audio-to-audio reference editing.
- Extreme Cost-Efficiency: The official cost per second at 2K resolution is less than 1/3 of mainstream models, and 768P is priced at half of what competitors charge for 720P.
- Technical Innovations: Utilizes H3-VAE tokenizer (4x sequence length efficiency), H3-Omni Transformer unified architecture, and heterogeneous compute optimization (30% training throughput boost). Training data consists entirely of real natural data.
Limitations: The H3-Regenerate-2K module and sparse attention implementation are not open-sourced initially, and detail precision in some scenarios still needs improvement.
Related event: MiniMax open-sources 33B multimodal video model H3(35 posts)→
More from Models
- Why are sub-15B open models so rare now? Labs seem to have settled on 24-27B minimum — Hot_Example_4456 · 2026-10-16
- User spots model openly generating "ideal responses" to train other models — liminal_bardo · 2026-10-16
- Reflection's 501B MoE Beam hits 463 tok/s on 8x GB200 with NVFP4 at day 0 — zhyncs42 · 2026-10-16
- Gemma 4 31B QAT on one 24GB card: 46 tok/s plain, ~80 with draft model on RTX 4090 — Subject_Poet_3861 · 2026-10-16
- JevBench v1.7.46 leaderboard: new d1-ultra tops 32 API systems at 76.1 composite — airesearch12 · 2026-10-16
- Team of 5 mulls dropping ChatGPT Pro for Business as OpenAI quota parity stays unclear — bettercallfool · 2026-10-16