MiniMax Is Building a Model Six Times Bigger Than Its Current Flagship
Chinese AI company MiniMax Group is working on a large language model built with 2.7 trillion parameters, and the company is reportedly planning to release it as open-weight software as soon as this quarter. If that timeline holds, it would be the largest open-weight model ever put out by a Chinese company — and possibly the largest open-weight model anywhere in the world.
The project carries the internal codename M3 Pro. That name places it as a successor to MiniMax's current flagship model, M3, which runs on 428 billion parameters. M3 Pro would be roughly six times larger than that existing system.
Where the Report Came From
A person with direct knowledge of MiniMax's plans told Reuters that the release could land in the third quarter of 2026. MiniMax itself declined to comment on the matter. The Information had reported on the development earlier, on July 8.
The company hasn't settled on a final release name for the model yet, so "M3 Pro" may end up being an internal placeholder rather than what eventually ships.
How M3 Pro Would Stack Up Against Existing Open-Weight Models
If MiniMax follows through, M3 Pro would leapfrog every trillion-parameter open-weight model currently available, including DeepSeek-V4-Pro, which sits at 1.6 trillion parameters and has held its position as one of the largest open releases to date.
MiniMax isn't new to building large, freely available models. Its existing MiniMax-M1 model already holds a notable distinction: the company describes it as the world's first open-source large-scale hybrid-attention reasoning model. M1 runs on 456 billion total parameters, supports a one-million-token context window, and was released under the Apache 2.0 license — meaning developers can use, modify, and build on it with minimal restriction.
What stands out about M1 isn't just its capability but its efficiency. MiniMax has reported achieving strong benchmark results while training the model for just $534,700 — a fraction of what comparable systems typically cost to train. That efficiency track record suggests M3 Pro may be built with the same emphasis on getting more performance out of less compute, even at a much larger parameter count.
Chinese AI Labs Are Racing to Release Bigger Open Models
MiniMax's plans aren't happening in isolation. They're part of a broader pattern among Chinese AI labs, which have been releasing increasingly capable models under permissive open licenses at a fast clip.
Tencent, for example, officially launched its Hy3 model on July 6. Hy3 is a mixture-of-experts model with 295 billion total parameters, though only 21 billion are active at any given time, and it was also released under Apache 2.0. Tencent says Hy3 performs on par with models carrying two to five times its parameter count, putting up reasoning benchmark scores that rival GLM-5.2 and DeepSeek-V4-Pro, while beating GPT-5.5 on scientific research tasks. Tencent has already folded Hy3 into its own product ecosystem, including WorkBuddy, CodeBuddy, Yuanbao, and QQ, and the company's Hong Kong-listed shares climbed more than 4% the day the model was announced.
This steady drumbeat of releases from Chinese labs — MiniMax, Tencent, and others including DeepSeek, Alibaba's Qwen, and Zhipu — is putting pressure on Western AI companies, many of which have moved toward closed or more tightly licensed releases for their most capable models. MiniMax, which is publicly listed in Hong Kong, is also planning to launch a separate frontier-level multimodal video generation model called H3 later this month, according to the same source who spoke to Reuters about M3 Pro.
Chinese Open-Weight Models Are Gaining Real Ground
The competitive shift shows up in usage data, too. According to JPMorgan strategist Michael Cembalest, Chinese models accounted for more than 45% of all traffic on OpenRouter by April 2026 — up from under 2% in late 2024. Part of the appeal is cost: JPMorgan's data indicates these open-source Chinese models run 60% to 90% cheaper than leading offerings from American labs like Anthropic and OpenAI, while delivering performance that comes close to matching the top end of the market.
Compute Limits Are Pushing Chinese Labs Toward Independence
U.S. export controls have restricted Chinese AI companies' access to the newest Nvidia hardware, and that's created what investors and executives have called a computing crunch across the industry. Some labs are responding by looking for alternatives outside American chip supply chains entirely. Z.ai, for instance, is reportedly exploring a custom AI chip as demand for its GLM models strains available compute. The company already showed it could operate independently of American silicon back in January, when it trained its GLM-Image model entirely on Huawei's Ascend chips.
Meanwhile, on the American side, Meta has said its next AI model will match OpenAI's GPT-5.5. The company expects to spend between $125 billion and $145 billion this year alone on chips, data centers, and other infrastructure. As part of that push, Mark Zuckerberg appointed Alexandr Wang, the former Scale AI founder, to lead Meta Superintelligence Labs in mid-2025, reportedly offering top researchers compensation packages worth hundreds of millions of dollars to join.
What Happens If M3 Pro Actually Ships
A model of this size raises real questions that go beyond the headline parameter count. Running 2.7 trillion parameters takes serious computational infrastructure, and it's not yet clear how accessible M3 Pro would be to developers and researchers outside of major, well-funded labs — even with open weights.
MiniMax's history with M1 offers a clue about how the company might approach this. Given that M1 delivered competitive results on a training budget under $600,000, MiniMax appears likely to lean on inference efficiency to make M3 Pro usable beyond just the handful of organizations with massive compute budgets. Whether that holds true at 2.7 trillion parameters, though, remains to be seen — MiniMax hasn't detailed how the model will be served or what hardware it will require to run.