MiniMax H3 generates a 10.125-second audiovisual file in under nine seconds
MiniMax's open-weight H3 now runs through vLLM-Omni, while its community license excludes the US, EU, UK and Republic of Korea.
By RuntimeWire Staff · Published
Primary source: MiniMax
Why it matters
FastH3 gives developers a public path to serve synchronized video and audio faster than playback, moving open-weight generation closer to responsive applications. MiniMax's standard license still requires separate permission for use in the US, EU, UK and Republic of Korea.
MiniMax, founded by Yan Junjie, now has a public serving path for its open-weight H3 model through vLLM-Omni and FastH3. MiniMax H3 generates a 10.125-second audiovisual file in under nine seconds on eight B300 GPUs, according to vLLM's benchmark.
The available test shows that a completed file can be produced faster than its playback duration. It does not establish first-frame latency or continuous streaming.
MiniMax released H3 under a community license in August 2026. MiniMax's H3 release announcement identifies it as an audiovisual generation model.
MiniMax called the stack "all public" in an official post on X, arguing that real-time generation makes interactive video possible.
The vLLM team used a narrower definition. Its completed MP4 file arrived faster than the file's playback duration. The test did not establish that frames can be shown as they are generated, which is the capability an interactive game, live agent or responsive world model would need.
Yan built MiniMax after spending more than six years at SenseTime, where he became a vice president and vice-head of its research institute. He studied mathematics at Southeast University, earned a doctorate in artificial intelligence from the Chinese Academy of Sciences and completed postdoctoral research at Tsinghua University. MiniMax's 2025 annual report credits him with roughly 200 academic papers and more than 30,000 citations.
Four denoising calls, plus the rest of the system
The speedup starts with FastH3, a distilled version of H3 developed through FastVideo with Hao AI Lab, Nuva Lab and Nvidia's FastGen team. FastH3 cuts H3's denoising process from 49 evaluations to four DiT calls.
That reduction alone does not produce a finished video. H3 first encodes prompts and reference media, generates video and audio latents together, decodes them through separate video and audio autoencoders, moves hundreds of frames off the GPUs and packages the result as an H.264 and AAC MP4, according to the vLLM technical report.
vLLM-Omni optimized that full path, including attention operations, communication between GPUs, parallel video decoding, output transport and MP4 construction. The published MiniMax H3 serving recipe exposes video generation through a familiar API service instead of leaving developers to assemble the pipeline themselves.
In the September 1 vLLM benchmark, FastH3 generated five-second, 10-second and 15-second clips faster than their respective playback durations. The 10.125-second output comprised 243 frames and completed in 8.678 to 8.710 seconds on eight Nvidia B300 GPUs. The 15-second request completed in 14.059 to 14.177 seconds. Benchmark outputs were generated at 1,344 by 768 pixels and 24 frames per second, with H.264 video and stereo 32 kHz AAC audio.
Hao AI Lab separately reported that FastH3 generated a 15-second, 768p clip in under 13 seconds on eight B200 GPUs and achieved up to a 14x speedup on one Blackwell GPU. That 14x figure remains a project-reported result. The vLLM team declined to calculate a base-H3-to-FastH3 speedup from its own experiments because the two test lanes used different code revisions, prompts, seeds and artifacts.
The initial FastH3 release is also specialized. It supports text-to-video-and-audio generation, while MiniMax's base H3 checkpoints also support first-and-last-frame conditioning and mixed image, video and audio references. The vLLM implementation fuses FastH3 into the model through load_weights() when the service loads; it does not operate as a request-level, switchable adapter.
The September 1 vLLM report said a stable raw benchmark bundle containing logs, environment details, media metadata and timing samples had not been published. A matched evaluation of base H3 and FastH3 across multiple seeds and output quality levels also remains on the project's work list.
"All public" has geographic and technical limits
MiniMax's H3 release included the base model checkpoints, encoders, tokenizers and audio and video decoders. Developers can run 768p generation locally and use public recipes for text, frame-conditioned and reference-guided audiovisual generation.
Two components described in MiniMax's H3 announcement are not part of the public serving stack. H3-Context-IR processes and expands multimodal input into a prompt for H3-Base, while H3-Regenerate-2K uses the base output, expanded prompt and original context to produce 2K video. Both modules remain accessible through MiniMax's hosted system.
The standard grant in the MiniMax H3 Community License excludes the European Union, United Kingdom, Republic of Korea and United States of America, requiring developers in those territories to obtain separate permission.
H3 is an open-weight base model with public inference artifacts and serving code.
MiniMax is turning research into distribution
The open serving work gives Yan a route to place H3 inside developer infrastructure that MiniMax does not control. Outside teams can optimize kernels, add hardware targets, produce quantized builds and adapt the model to production workflows. vLLM and Hao AI Lab gain a prominent audiovisual workload for their serving and acceleration projects. Nvidia gets another benchmark that puts eight Blackwell GPUs at the center of the result.
MiniMax has a financial reason to welcome outside optimization. Its 2025 annual report reported $79 million in revenue in 2025, up from $30.5 million in 2024. MiniMax also reported more than 236 million cumulative users, 214,000 enterprise customers and developers, and more than 600 million videos generated through the end of 2025. Those are MiniMax's figures, and cumulative reach does not show how many people or organizations pay.
A model that other engineers can serve more cheaply has a better chance of becoming infrastructure instead of remaining a feature inside MiniMax's Hailuo video product. The territorial license limits that distribution in several of the markets most likely to fund and commercialize interactive video applications.
The H3 work also extends MiniMax's recent campaign around open weights. In an August 13 announcement, MiniMax described Music 3.0 as an open-weight model for generating songs up to five minutes long using an eight-layer residual vector quantization design and a global-local language model.
FastH3 is the more concrete release. The weights, pre-extracted LoRA and serving code give developers something they can inspect and improve today, where permitted. The result is a meaningful systems milestone for completed audiovisual generation. Interactive video will require the next set of numbers: time to first frame, continuous streaming latency, concurrent-request performance, cost and output quality under the same test conditions.