Moonshot's Kimi K3: the largest open-weight model ever, benchmarking at the US frontier

Moonshot AI unveiled Kimi K3, a 2.8-trillion-parameter model that becomes the largest open-weight AI system ever released — the first to approach the three-trillion mark — with a 1M-token context window and text-plus-image input, built for long-horizon reasoning and coding. The early independent signal is striking: in blind testing on Arena, developers preferred K3 over every leading US model for front-end coding, and in the broader text ranking it tied OpenAI's flagship while outranking models that sat at the frontier weeks ago. The weights themselves are scheduled for release on July 27; until then, no one outside Moonshot can inspect or run it.

Why it matters

This is the strongest evidence yet that China's open-weight ecosystem isn't months behind the US frontier — it's inside benchmark noise of it, and giving the result away. For the closed US labs the squeeze is now two-sided: Thinking Machines' Inkling attacks from the customization flank while K3 attacks on raw capability at open-weight prices, on top of an existing API price war. And for China's AI capital markets — DeepSeek's $74B round-and-IPO track — a domestic lab benchmarking at the frontier is the narrative fuel that keeps valuations compounding.

What to watch

Whether the July 27 weights release actually ships at full scale (a 2.8T-parameter download is its own logistics problem), independent replication of the Arena results once outsiders can run it, and pricing of Moonshot's hosted API against US flagships.