SGLang
The SGLang cookbook documents multi-GPU serving for MiniMax-H3.
lmsysorg.mintlify.appRetrieved
Open models
This reference covers the open base checkpoint for an omni-modal video generation system, with higher-resolution modules remaining hosted. Its community license excludes the European Union, the United Kingdom, the Republic of Korea, and the United States of America.
MiniMax H3 is the vendor reference for MiniMaxAI/MiniMax-H3, as of Aug 12, 2026. Omni Transformer: H3-Omni-Transformer is a 33B-parameter dense, single-stream Transformer, with approximately 13B parameters residing in AdaLN-related branches.; Output duration: 4-15 seconds, as of Aug 12, 2026. RunInfra has not measured this model; every figure below belongs to its named source, cited and dated.
| Group | Fact | Source-cited display value | Source and date |
|---|---|---|---|
| Identity | Release identity | MiniMaxAI/MiniMax-H3 | huggingface.coRetrieved |
| Identity | Vendor description | MiniMax H3 is a general-purpose, omni-modal generative system. It supports unified understanding of multimodal contexts composed of text, images, video, and audio, and can generate video with native stereo audio at resolutions up to 2K and durations of up to 15 seconds. | huggingface.coRetrieved |
| Identity | Launch publication date | 2026-07-31 | www.minimax.ioRetrieved |
| Identity | Open-source publication date | 2026-08-03 | www.minimax.ioRetrieved |
| Identity | License-stated date | MiniMax H3 release date/License date: August 2, 2026. | huggingface.coRetrieved |
| Architecture | Omni Transformer | H3-Omni-Transformer is a 33B-parameter dense, single-stream Transformer, with approximately 13B parameters residing in AdaLN-related branches. | huggingface.coRetrieved |
| Architecture | AdaLN deployment | AdaLN-related branches are cacheable for inference-only deployment. | huggingface.coRetrieved |
| Architecture | Position encoding | 3D MM-RoPE over (t, h, w) | huggingface.coRetrieved |
| Architecture | Text and vision encoder | uses the full pretrained weights of Qwen3-VL-32B; layer-50 hidden states | huggingface.coRetrieved |
| Architecture | Video VAE | f16t4d24; temporally causal | huggingface.coRetrieved |
| Architecture | Audio VAE | 32 kHz stereo to 40 Hz latents per channel | huggingface.coRetrieved |
| Architecture | Joint prediction | joint audio and video prediction in one transformer | huggingface.coRetrieved |
| Architecture | Initial attention path | The initial open-source release provides inference with full attention only. | huggingface.coRetrieved |
| Architecture | Released checkpoints | The released checkpoints are CFG-distilled Omni Transformer model weights. | huggingface.coRetrieved |
| Context | Output duration | 4-15 seconds | huggingface.coRetrieved |
| Context | Output frame rate | 24 FPS | huggingface.coRetrieved |
| Context | Output audio | 32 kHz stereo | huggingface.coRetrieved |
| Context | Output resolution | Shorter side 768 pixels by default; 2K generation can be achieved with H3-Regenerate-2K. | huggingface.coRetrieved |
| Context | Aspect ratios | including but not limited to 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16 | huggingface.coRetrieved |
| Context | Languages | Stable support for 11 languages | huggingface.coRetrieved |
| Modalities | FL2VA checkpoint | FL2VA supports t2va and fl2va with text plus optional first and last frames; output is video and audio. | huggingface.coRetrieved |
| Modalities | Ref2VA checkpoint | Ref2VA accepts text with up to 9 reference images, up to 3 videos, up to 3 audio clips, and a maximum of 12 files; output is video and audio. | huggingface.coRetrieved |
| License | License identity | MiniMax H3 COMMUNITY LICENSE AGREEMENT | huggingface.coRetrieved |
| License | Excluded territories | 'Excluded Territories' means the European Union, the United Kingdom, the Republic of Korea and the United States of America. | huggingface.coRetrieved |
| License | Territorial output restriction | You may not use, reproduce, modify, distribute, or display the MiniMax H3 Works or any of their Outputs or results outside the Applicable Territory. | huggingface.coRetrieved |
| License | Revenue clause | Separate authorization is required if your commercial products and services generate more than 20 million US dollars (or equivalent in other currencies) in yearly revenue. | huggingface.coRetrieved |
| License | Output ownership | MiniMax claims no rights over the Outputs you generate. | huggingface.coRetrieved |
| License | Encoder license | the encoder of MiniMax H3 uses Qwen3-VL-32B, which is licensed under Apache 2.0 License | huggingface.coRetrieved |
| License | Excluded-territory application | https://platform.minimax.io/h3-license | huggingface.coRetrieved |
| Pricing | MiniMax API price: Vendor price comparison | At 2K, H3's per-second price is less than a third of mainstream models | www.minimax.ioRetrieved |
| Availability | Repository access | Ungated repository. | huggingface.coRetrieved |
| Availability | Repository formats | Original FL2VA/ and Ref2VA/ formats and diffusers format are provided side by side in one repository. | huggingface.coRetrieved |
| Availability | Open base module | H3-Base: Generates audio and video based on the H3-Context-IR output, producing results at 768p resolution. | huggingface.coRetrieved |
| Availability | Context module | H3-Context-IR is not included in this open-source release. | huggingface.coRetrieved |
| Availability | Higher-resolution module | H3-Regenerate-2K is not yet open-sourced. | huggingface.coRetrieved |
| Availability | Diffusers integration | Modular Diffusers blocks integration was merged on August 5, 2026. | github.comRetrieved |
| Availability | ComfyUI support | Native support requires version 0.30.0 or later. | docs.comfy.orgRetrieved |
| Availability | Task checkpoint weights | MiniMaxAI/MiniMax-H3 | huggingface.coRetrieved |
"Today, we're officially launching MiniMax H3, a general-purpose omni-modal generation model. H3 can jointly understand multimodal contexts spanning text, images, video, and audio. It generates video with native stereo audio at up to 2K resolution and 15 seconds in length."
www.minimax.ioRetrieved
"Today, we are officially open-sourcing MiniMax H3, our next-generation general-purpose video model."
www.minimax.ioRetrieved
These results are vendor-claimed, not independently measured by RunInfra.
No vendor-claimed benchmark results for this exact build were present in the cited material as of Aug 12, 2026.
Listed rows have a cited upstream support signal. They are not RunInfra measurements.
The SGLang cookbook documents multi-GPU serving for MiniMax-H3.
The official vLLM recipes site publishes a MiniMax-H3 recipe linked from the model card.
RunInfra has not measured this model yet. When measurement is published, the record will state throughput, latency, memory, quality, serving conditions, and reproducible evidence.
As of Aug 12, 2026, Release identity: MiniMaxAI/MiniMax-H3.
As of Aug 12, 2026, Repository access: Ungated repository.
As of Aug 12, 2026, Repository access: Ungated repository.
Use a workspace API key and pay for input, cached input, and output tokens.
View Model APIs