# MiniMax puts four models on GMI Cloud's two-week unlimited tab > Source: > Published: 2026-08-24 18:13:33+00:00 # MiniMax puts four models on GMI Cloud's two-week unlimited tab **MiniMax, founded by Yan Junjie, is offering M3, M2.7, Speech 2.8 and Music 3.0 through GMI Cloud, while its usage limits and Music 3.0 eligibility remain unclear.** By [RuntimeWire Staff](/author/runtimewire-staff) ยท Published Primary source: [MiniMax](https://x.com/MiniMax_AI/status/2091948930124947941) ## Why it matters The promotion gives GMI Cloud users access to four named MiniMax models, a subset of the company's broader language, video, audio and music catalog. In an [August 24 post](https://x.com/MiniMax_AI/status/2091948930124947941?ref=runtimewire), [MiniMax](https://platform.minimax.io/?ref=runtimewire) said developers could get 14 days of what it called unlimited access to four models through [GMI Cloud](https://www.gmicloud.ai/en/company/about?ref=runtimewire). [MiniMax](/models/fal/minimax-preview-speech-2.5-hd) was founded by [Yan Junjie](https://ir.minimax.io/corporate-information/management?ref=runtimewire). The promotion runs through September 6 and includes [MiniMax M3](/models/minimax/minimax-m3), M2.7, Speech 2.8 and Music 3.0. It puts coding, agentic workflows, voice generation and music generation in one trial. Yan, MiniMax's founder, chairman, CEO and CTO, came to foundation models through computer vision. He spent more than six years at SenseTime, rising to vice president and vice-head of its research institute, after studying mathematics at Southeast University, earning an artificial-intelligence doctorate from the Chinese Academy of Sciences and conducting postdoctoral research at Tsinghua University. MiniMax says Yan has published about 200 academic papers with over 30,000 citations. That research career sits behind a distinctly product-driven thesis. Yan has argued that AI should reach users as a practical product and move beyond research settings. MiniMax's [full-year results](https://www.minimax.io/news/minimax-global-announces-full-year-2025-financial-results?ref=runtimewire) made the commercial version of that ambition explicit: Yan said MiniMax planned to "evolve from a large-model company into a platform company for the AI era." The GMI Cloud promotion is a small but concrete step in that direction. MiniMax gets another route to developers, while GMI Cloud gets recognizable models that can bring developers into its infrastructure. ### Four models, one acquisition funnel MiniMax [describes M3](https://www.minimax.io/blog/minimax-m3?ref=runtimewire) as a multimodal coding and agentic model built around MiniMax Sparse Attention. MiniMax's [model documentation](https://platform.minimax.io/docs/guides/models-intro?ref=runtimewire) lists a context window of up to 1 million tokens, alongside image and video input. MiniMax says M2.7 targets engineering, office work and recursive-improvement workflows. The audio side broadens the trial beyond the crowded coding-model market. MiniMax lists HD and Turbo versions of Speech 2.8, with support for 40 languages and seven emotions. MiniMax's [Music 3.0 documentation](https://platform.minimax.io/docs/api-reference/music-generation?ref=runtimewire) describes the model as supporting intent understanding, sound enhancement and more humanized vocals. Bundling those models makes the promotion useful as a product-development test rather than a single-model benchmark run. A developer can use M3 or M2.7 for application logic, add synthesized speech, and experiment with generated music within the same provider relationship. That is the pitch GMI Cloud needs as GPU providers compete to sell higher-level inference services instead of interchangeable access to chips. GMI Cloud pairs MiniMax model access with its GPU infrastructure. Its current [company materials](https://www.gmicloud.ai/en/company/about?ref=runtimewire) list regions in North America, Europe and Asia-Pacific. In its own [pricing comparisons](https://www.gmicloud.ai/en/blog/runpod-lambda-coreweave-gmi-gpu-rental-pricing?ref=runtimewire), the California-based provider compares its offering with specialist AI clouds including [CoreWeave](https://www.coreweave.com/news/coreweave-to-acquire-monolith-expanding-ai-cloud-platform-into-industrial-innovation?ref=runtimewire), [RunPod](https://runpod.io/?ref=runtimewire) and [Lambda](https://lambda.ai/?ref=runtimewire). A separate [inference-platform comparison](https://www.gmicloud.ai/en/blog/choosing-an-llm-inference-platform-gmi-cloud-together-ai-and-the-production-ai-stack-compared?ref=runtimewire) sets GMI Cloud against [Together AI](https://www.together.ai/blog/announcing-our-series-c?ref=runtimewire). Each provider is trying to own the layer between model builders and the developers paying to run inference. ### The word "unlimited" is doing considerable work MiniMax and GMI Cloud have not attached public rate limits, concurrency rules, qualifying account types or fair-use terms to the posts announcing the promotion. "Unlimited" therefore describes the offer's billing pitch, while saying little about the throughput a developer can expect during the 14-day window. MiniMax's [music-generation documentation](https://platform.minimax.io/docs/api-reference/music-generation?ref=runtimewire) says that, beginning August 20, its paid music-generation and lyrics-generation APIs would no longer accept new users. The same notice says MiniMax would discontinue its free music APIs and direct users toward MiniMax Audio or the open-weight [Music 3](/models/fal/music-3) model. The GMI Cloud offer started four days after that cutoff and explicitly includes Music 3.0. A partner-hosted route could sit outside MiniMax's restrictions for new users on its own API, but the promotion does not explain the distinction. Developers considering the offer will need to establish whether Music 3.0 access covers new GMI Cloud accounts, which regions qualify and what happens after September 6. That ambiguity does not erase the value of the trial. It sets the practical test MiniMax and GMI Cloud must pass: developers need enough capacity to build something substantial, rather than a brief demo that reaches an undisclosed ceiling. ### GMI Cloud needs models to fill the machines GMI Cloud founder and CEO Alex Yeh originally built infrastructure for Bitcoin computing nodes before shifting toward AI GPUs as customer demand changed, he [told TechCrunch](https://techcrunch.com/2024/10/29/gmi-cloud-secures-82m-in-series-a-for-its-gpu-cloud-infrastructure/?ref=runtimewire). In 2024, TechCrunch reported that GMI Cloud raised an $82 million Series A led by Headline Asia, with Banpu Next and Wistron participating. The financing consisted of $15 million in equity and $67 million in debt. In July 2026, GMI Cloud [committed $500 million in capital spending](https://www.gmicloud.ai/en/blog/gmi-cloud-commits-500-million-to-expand-ai-infrastructure-for-frontier-ai-customers?ref=runtimewire) to expand its computing capacity. GMI Cloud also said it had secured nine-figure contracts with an unnamed US frontier AI customer. The MiniMax promotion follows that infrastructure commitment, turning an infrastructure expansion story into a developer-acquisition campaign. MiniMax brings the model catalog; GMI Cloud supplies the compute and customer account. Both sides benefit if applications created during the promotion continue consuming inference after September 6. MiniMax has spent August putting more of its model portfolio in front of US developers. RuntimeWire previously reported on its [M3 inference teaser with SambaNova](/article/minimax-m3-sambanova-inference-speed-test-teaser) and a series of [San Francisco events around the H3 video model](/article/minimax-h3-modcon-morgan-suo-us-developer-push). The GMI Cloud offer is more actionable than a performance teaser because it names the models, access provider and trial dates. Its unanswered usage terms will determine how much developers can actually build. For Yan, that is the point of the exercise. MiniMax's platform ambitions depend on developers encountering its models inside the tools and clouds they already use. Retaining those developers will depend on the limits, reliability and pricing waiting behind the promotion.