Three Chinese Labs Cross 2T Parameters in Open Weights Within Weeks三家中国实验室数周内相继推出逾2万亿参数开源模型

Within a single week centered on WAIC 2026 in Shanghai, three Chinese AI labs staked claims above the two-trillion-parameter threshold for open-weight models. Moonshot shipped Kimi K3 at 2.8 trillion parameters; Alibaba previewed Qwen3.8-Max at 2.4 trillion; and The Information reported that MiniMax is preparing an M3 Pro at 2.7 trillion for a potential Q3 open-source release — roughly six times larger than its current M3.

The concentration is notable because all three are framed as open weights, not API-only products. MiniMax's current M3, which shipped June 1 at 428 billion parameters, already priced input at $0.12 per million tokens versus $5 for comparable closed models. M3 Pro at 2.7 trillion would dwarf it — though MiniMax has confirmed nothing officially.

For context: Qwen now anchors more than 200,000 derivative models on Hugging Face and has surpassed one billion cumulative downloads. The open-weight race has material infrastructure implications for anyone building on these model families globally.