MiniMax Releases H3, a 33B-Parameter Omni-Modal Model That Generates 2K Video With Native Stereo Audio
MiniMax has published MiniMax-H3, a 33-billion-parameter omni-modal generative model capable of producing up to 15 seconds of 2K video with native stereo audio. The model accepts text, image, video, and audio inputs, though its full 2K pipeline depends on a hosted preprocessing component not included in the open-source release.
MiniMax Releases H3, a 33B-Parameter Omni-Modal Video and Audio Generation Model
MiniMax has released MiniMax-H3, a 33-billion-parameter dense transformer that generates video with native synchronized stereo audio, according to the model card published on Hugging Face. The system accepts text, image, video, and audio as input and outputs video up to 2K resolution and 15 seconds in duration at 24 FPS with 32 kHz stereo audio.
H3 is built around three components: H3-Context-IR, a hosted preprocessing system that converts free-form multimodal instructions into a structured intermediate representation; H3-Base, which generates 768p video and audio from that representation; and H3-Regenerate-2K, which upsamples the 768p output to 2K by feeding it back through the model alongside the original context. Notably, H3-Context-IR itself is not included in this open-source release — MiniMax says it relies on a multi-stage workflow across multiple hosted models and is instead exposed only through an API, with a "Prompting Guidance" document for developers who want to replicate its behavior locally.
Architecture Details
According to MiniMax, the core H3-Omni-Transformer is a 33B-parameter dense, single-stream design, with roughly 13B parameters sitting in AdaLN-related branches that can be precomputed and cached, reducing the load footprint for inference-only deployments. Text, image, and video inputs are encoded through an H3-Encoder built on the full pretrained weights of Qwen3-VL-32B, drawing hidden states from its 50th layer. Video and audio are separately compressed via H3-VisualVAE (16x spatial compression, 4x temporal compression, 24 latent channels) and H3-AudioVAE (32 kHz stereo audio compressed to a 40 Hz latent token rate).
The model natively supports sparse-attention training and inference for handling long multimodal sequences, but MiniMax states that this initial release only supports full-attention inference — the sparse-attention implementation is planned for a future update, with no timeline disclosed.
Input Modes and Language Support
Two model variants are available: H3-Base-FL2VA, which supports zero, one, or two input images for text-to-video, first-frame, last-frame, or first-and-last-frame generation; and H3-Base-Ref2VA, an "omni-reference" mode accepting up to 9 images, 3 video clips, and 3 audio clips (audio cannot be the sole input) with a combined cap of 12 files. The model supports 11 languages with stable dialogue generation — Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish — plus unspecified partial support for additional languages.
H3 is accessible via API at platform.minimax.io (global) and platform.minimaxi.com (China), and through consumer apps including the Hailuo AI web app and MiniMax Hub desktop client. Pricing for API access was not disclosed in the model card.
MiniMax has not published standardized benchmark scores (e.g., VBench or similar video-generation metrics) alongside this release, and the model is distributed under a MiniMax H3 Community License rather than a fully permissive open-source license, with content moderation guardrails applied to user inputs and outputs.
What this means
H3 pushes MiniMax further into full-stack multimodal generation, competing with video models from Alibaba, ByteDance, and Kuaishou that also target synchronized audio-video output. The decision to withhold H3-Context-IR — the component responsible for parsing complex multimodal instructions — means the open-weight release is not a complete replacement for MiniMax's hosted pipeline; developers wanting the full 2K quality path will still depend on MiniMax's API or need to reverse-engineer the preprocessing behavior themselves. The 33B parameter count with heavy AdaLN caching optimizations suggests MiniMax is prioritizing deployability, but until independent benchmarks and pricing are available, comparisons to closed competitors like Runway, Kling, or Veo remain speculative.
Related Articles
Anonymous Provider Launches Union Alpha, a Free 262K-Context Multimodal Model on OpenRouter
A third-party provider using the alias 'Stealth' has released Union Alpha on OpenRouter, a multimodal model with a 262K context window, currently free to use during its preview period. The model's developer remains anonymous, and OpenRouter states it is not the model's owner or operator.
DeepSeek Ships V4.1-Flash With Novel Encoder-Decoder Architecture, Cuts KV Cache to 1/8 of Predecessor
DeepSeek released V4.1-Flash, a 763B-parameter model built on a new causal encoder-decoder architecture that splits 8B active parameters for prefill and 16B for decode. The model adds native vision support, a 1M-token context window, and shrinks KV cache footprint to roughly 1/8 of DeepSeek V4 Flash, while retiring V4 Pro.
InclusionAI Releases Ling 3.0 Flash VL, Adding Vision to Its 124B MoE Model
InclusionAI has released Ling 3.0 Flash VL, a vision-language extension of its 124B total-parameter, 5.5B active Mixture-of-Experts model. The model adds native image and video understanding, supports a 131K token context window, and is priced at $0.06 per 1M input tokens and $0.18 per 1M output tokens via OpenRouter.
Ex-OpenAI Researcher Launches Jev, an AI Model That Scores Options Instead of Generating Text
Startup TypeSafe AI has released Jev, a model built to score predefined answer options rather than generate text, claiming response times of 70 to 500 milliseconds. Co-founder Diogo Almeida, a former OpenAI researcher and InstructGPT co-author, says the model targets background classification tasks like sorting customer requests.
Comments
Loading...