MiniMax brings H3 video generation to Luma Agents, with access terms unclear

Why it matters

Availability through Luma Agents could place MiniMax's H3 inside a collaborative production workflow alongside competing video models. Pricing, eligibility, regional availability, usage limits and developer access remain undocumented.

Luma said on August 6, 2026, that MiniMax's H3 video model was available in Luma Agents, its multi-model creative workflow product. MiniMax, founded by Yan Junjie, separately amplified the announcement.

Luma's announcement

MiniMax's official account said H3 could generate up to 15 seconds of 2K video with native stereo sound through Agents.

The announcements leave important commercial and technical details unresolved. Neither specifies pricing, eligibility, regional availability, usage limits or whether developers can invoke H3 through a Luma API.

MiniMax, based in Shanghai, was founded in early 2022 by Yan, according to MiniMax's official company profile. Luma is based in Palo Alto and is led by co-founder and CEO Amit Jain.

Where H3 fits in Luma Agents

Luma launched Agents on March 9 as a shared creative board where agents generate assets, manage revisions and retain project context. Its official overview says the product can route work across image, video and audio models and lets advanced users access individual models through the board interface.

Luma's published product material does not document H3 as a supported model. MiniMax's announcement therefore establishes the claimed availability, while Luma's materials do not explain how users select H3, whether an agent can route work to it automatically or what project context reaches the model.

MiniMax introduced H3 on July 31 as a general-purpose multimodal generation model that accepts text, image, video and audio context. MiniMax says it supports native stereo audio, output of up to 2K resolution and clips as long as 15 seconds. MiniMax also lists video-to-video motion transfer, multimodal editing and multi-shot generation among its capabilities. Its quality and price-performance assertions remain MiniMax claims.

Those specifications suit an application designed to coordinate several media types within one project. Availability through Agents could also expose H3 to customers working outside MiniMax's own Hailuo product and developer platform. The supplied materials do not disclose how many Luma users can access it or whether the integration is generally available.

A crowded market for video with sound

H3 arrives as video-model developers add synchronized speech, music and ambient sound to systems that previously centered on silent clips. Kuaishou's Kling 3.0 also supports native audio and clips of up to 15 seconds. Google's Veo combines generated video with dialogue, sound effects and ambient audio. ByteDance has positioned Seedance 2.0 around multimodal audio-video generation, while Runway continues to develop its own video and world-model systems.

For MiniMax, placement inside Agents offers distribution in a product that already coordinates competing model families. For Luma, H3 adds another option for projects requiring longer clips and generated sound, provided the announced access becomes available on terms customers can use.

The founders came from different technical paths

Yan is MiniMax's founder, chairman, CEO and CTO. According to MiniMax's management biography, he studied mathematics at Southeast University, earned a doctorate in artificial intelligence from the Institute of Automation at the Chinese Academy of Sciences and completed postdoctoral research at Tsinghua University. Before founding MiniMax, he spent more than six years at SenseTime and became vice president and deputy head of its research institute.

Jain approached creative AI from a systems background. His South Park Commons profile says he worked as a systems and machine-learning engineer at Apple, led work on Passthrough for Vision Pro and helped bring LiDAR to the iPhone. He began building Luma in 2021 after concluding that learned models would assume a larger role in creative production.

MiniMax and Luma have raised substantial capital

According to its investor-relations company profile, MiniMax says its proprietary models and AI-native products have cumulatively served more than 212 million individual users across more than 200 countries and regions, as well as more than 100,000 enterprises and developers across more than 100 countries and regions. These are company-reported figures. Its Hong Kong listing prospectus reported revenue of US$53.4 million for the nine months ended September 30, 2025, up from US$19.5 million in the corresponding 2024 period. MiniMax had 385 full-time employees at the end of that period, including 284 in research and development, according to the prospectus.

Alibaba led a private financing of at least $600 million in 2024 that valued MiniMax above $2.5 billion, Bloomberg reported. MiniMax's named backers have also included Tencent, MiHoYo, Hillhouse, HongShan, IDG Capital, the Abu Dhabi Investment Authority, Boyu Capital, Mirae Asset, Aspex and Eastspring. MiniMax subsequently raised $619 million in a January 2026 Hong Kong initial public offering, according to Reuters.

Luma announced a $900 million Series C led by Humain, with significant participation from AMD and backing from Andreessen Horowitz, Omniva, Amplify Partners and Matrix Partners. Luma said it had more than 130 employees at the time of that announcement, but Luma did not disclose revenue or total user figures in the supplied materials.

Access terms remain undocumented

The supplied materials do not establish whether H3 is available to every Agents customer, restricted to selected accounts or limited by region. They also do not disclose pricing, usage allowances or account requirements.

MiniMax's official documentation describes a video-generation API, but it does not establish whether developers can invoke H3 through Luma Agents. Until Luma documents the integration, the announcement supports a product-availability claim rather than a defined developer offering.