Chinese AI startup MiniMax is reportedly developing a 2.7 trillion-parameter model, internally codenamed M3 Pro, roughly six times the size of its current flagship M3, according to reporting from The Information and Reuters on July 8. MiniMax has not confirmed the model publicly, and no release date is official, though the reports point to an open-source launch as early as the third quarter.
Treat the parameter count as a headline, not the story. What it signals about the pace of China’s AI buildout matters more to business readers than the model itself, and it did not arrive alone.
What is actually new
MiniMax open-sourced its current M3 model, 428 billion parameters, in June. A 2.7 trillion-parameter successor within one quarter would be an aggressive scaling pace even by frontier-lab standards, and MiniMax has built its reputation on releasing open weights rather than gating access behind an API, a strategy that has made its models a default choice for developers who cannot or will not use a Western frontier lab’s hosted service.
The same week, Ant Group’s robotics arm Robbyant open-sourced LingBot-VLA 2.0, an upgraded vision-language-action model meant to serve as a general-purpose control system for robots, trained on roughly 60,000 hours of real-world physical data. And Chinese chipmaker MetaX has its C600 processor, built on HBM3e memory with FP8 support, moving toward full mass production this year, part of a broader push toward domestically produced AI silicon that does not depend on export-controlled Nvidia hardware.
What this means for procurement and risk planning
None of these three developments individually changes much for a business outside China. Together, they describe an ecosystem building full-stack independence, models, robotics infrastructure, and chips, on a timeline that does not wait for export control policy to loosen. We have covered the chip side of this before: DeepSeek is reportedly building its own inference chip, and labs increasingly see owning silicon as the only way to control their own inference costs. MiniMax, Robbyant, and MetaX are the same logic applied at the scale of an entire national AI stack rather than one company.
For procurement and risk teams at multinational firms, the practical question is not whether to use Chinese open-weight models today. It is whether vendor lists and AI tooling roadmaps account for a scenario where China-origin open models and China-origin chips become the fastest-improving, lowest-cost option in categories Western labs currently lead. Open-weight strategy is itself a form of the lock-in economics we have tracked elsewhere: MiniMax giving away weights for free is a customer-acquisition move, not a act of altruism, and it works precisely because it undercuts the pricing power of closed, API-gated competitors.
What to watch
Watch for MiniMax’s own confirmation, since everything here is currently sourced to reporting rather than an official announcement, and the actual model, if it ships, may differ from what has leaked. Watch too for whether Western labs respond to open-weight pressure from Chinese competitors with more aggressive open releases of their own, the clearest sign this dynamic is reshaping strategy rather than just headlines. The clearest case study already exists: Moonshot’s Kimi is the open Chinese model already inside Western coding tools.
AI and business tech news, verified by a physician who reads the filings. One email a week, no noise.
