MiniMax launches open-weight H3 and claims video benchmark lead

MiniMax, founded by Yan Junjie, launched its open-weight H3 video-generation model on July 31 and claims it is the highest-ranked open-weight video model on both Arena and Artificial Analysis. However, the Artificial Analysis leaderboard lists H3 at No. 2 overall behind Google Gemini Omni Flash, and the Arena leaderboard did not list H3 as of Aug. 3, making MiniMax's Arena claim unverified. H3 supports text-to-video, image-to-video, first-last-frame generation, and multimodal reference inputs, with native stereo audio, and is available via open weights on Hugging Face and ComfyUI, as well as commercial API.

MiniMax https://platform.minimax.io/?ref=runtimewire , founded by Yan Junjie https://ir.minimax.io/corporate-information/management?ref=runtimewire , said days after launching H3 that the model had become the highest-ranked open-weight video model on both Arena https://arena.ai/leaderboard/text-to-video?rankBy=labs&ref=runtimewire and Artificial Analysis https://artificialanalysis.ai/embed/text-to-video-leaderboard/leaderboard/text-to-video?ref=runtimewire . MiniMax on X https://x.com/MiniMax AI/status/2084410437618352386?ref=runtimewire MiniMax announced H3 https://www.minimax.io/blog/minimax-h3?ref=runtimewire on July 31. Yan, MiniMax's chairman, CEO and CTO, spent over 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 later conducted postdoctoral research at Tsinghua University. MiniMax's prospectus says he has published about 200 academic papers with more than 30,000 citations. HKEX prospectus https://www1.hkexnews.hk/listedco/listconews/sehk/2025/1231/2025123100025.pdf?ref=runtimewire MiniMax's H3 blog frames the release as part of a push to "break task boundaries" across text, image, video and audio. The release's practical distinction is downloadable weights: the Comfy-Org Hugging Face model card https://huggingface.co/Comfy-Org/MiniMax-H3?ref=runtimewire provides weight-download instructions, and ComfyUI's H3 documentation https://docs.comfy.org/tutorials/video/minimax/minimax-h3?ref=runtimewire gives developers local workflow examples. The public evidence is narrower than MiniMax's claim MiniMax's benchmark post https://x.com/MiniMax AI/status/2084410437618352386?ref=runtimewire says H3 is the leading open video-generation model on Artificial Analysis. MiniMax made the same claim for Arena. The Artificial Analysis leaderboard reviewed for this story listed H3 at No. 2 overall behind Google Gemini Omni Flash and first among open-weight entries, with Elo 1,239 and 6,034 samples. The Arena side requires careful attribution. The live Arena text-to-video page https://arena.ai/leaderboard/text-to-video?rankBy=labs&ref=runtimewire reviewed during reporting on Aug. 3 did not list H3. MiniMax's Arena ranking therefore remains a MiniMax claim rather than a result independently visible on the public leaderboard. Arena on X https://twitter.com/arena/status/2084408459991421319?ref=runtimewire Both rankings rely on preference comparisons. They capture which outputs evaluators prefer in controlled matchups. They do not exhaustively test production work, including inference speed on specific hardware, consistency across long projects, controllability after deployment or legal exposure from generated content. Yan's bet is task generalization H3 supports text-to-video, image-to-video, first-last-frame generation and multimodal reference inputs, with native stereo dialogue, music and sound effects. MiniMax says H3 can generate clips of 4 to 15 seconds at 24 frames per second, with output at 768P or 2K. MiniMax says H3 can jointly understand multimodal contexts spanning text, images, video and audio, then generate video with native stereo audio. MiniMax's video-generation documentation https://platform.minimax.io/docs/guides/video-generation?ref=runtimewire lists H3 across text-to-video, image-to-video, first-last-frame and multimodal reference workflows. The open-weight release adds a distribution route unavailable to services that provide access only through proprietary interfaces. The Hugging Face release https://huggingface.co/Comfy-Org/MiniMax-H3?ref=runtimewire and ComfyUI's H3 documentation https://docs.comfy.org/tutorials/video/minimax/minimax-h3?ref=runtimewire give developers a path to run released H3 components locally. MiniMax also offers commercial API access, so open-weight experimentation can feed usage of its paid platform. A benchmark claim with financial stakes H3 is arriving as Yan pushes MiniMax from a research-led model maker toward a platform spanning consumer applications, developer APIs and downloadable weights. MiniMax's HKEX prospectus showed revenue of $30.5 million in 2024 and $53.4 million for the nine months ended Sept. 30, 2025, while adjusted net loss for that nine-month period was $186.3 million. Average monthly active users rose from 3.1 million in 2023 to 27.6 million for the nine months ended Sept. 30, 2025, and new AI-native product users rose from 11.1 million in 2023 to 104.2 million in 2024. HKEX prospectus https://www1.hkexnews.hk/listedco/listconews/sehk/2025/1231/2025123100025.pdf?ref=runtimewire MarketScreener reported in July that MiniMax was seeking HK$16.04 billion, or about $2.05 billion, through a share sale and convertible bonds to fund AI growth. That financing effort shows the capital demands around a market where Google, ByteDance, OpenAI, Alibaba, Kuaishou and independent labs are releasing video systems at a rapid pace. A leaderboard lead can attract developers during the release window; sustained adoption will depend on whether MiniMax can convert open-weight experimentation into API demand and finished products. MarketScreener https://www.marketscreener.com/news/china-s-minimax-seeks-to-raise-total-2-05-billion-via-share-sale-bond-issue-ce7f5ededc8eff20?ref=runtimewire MiniMax attached a claimed benchmark lead to H3 only days into its release, although the Arena evidence remains incomplete. Developers will determine whether H3's general-purpose design can hold characters, motion, sound and creative direction together across repeatable commercial workflows.