China's Kimi K3 Is Now the Biggest Open AI Model on Earth

China's Kimi K3 Is Now the Biggest Open AI Model on Earth

Somewhere in a Beijing office, a whiteboard just got a number so large it needs its own zip code: 2.8 trillion parameters. Moonshot AI's new Kimi K3 isn't just big for a Tuesday — it's the largest open-weight AI model anyone has shipped, and it's daring Western labs to match it in public.

The Numbers That Broke the Curve

Moonshot AI released Kimi K3 on July 16, a 2.8-trillion-parameter, open-weight, multimodal reasoning model with a 1-million-token context window and an always-on "thinking mode." Pricing lands at $3 per million input tokens and $15 per million output tokens — a fraction of what comparable Western frontier models charge. Full open weights are set to land by July 27.

Moonshot says the model posts frontier-level scores across a range of evaluations, claiming it consistently outperforms rivals including OpenAI's GPT-5.5, Anthropic's Claude Opus 4.8, and China's own Zhipu GLM-5.2 on the benchmarks it chose to publish. Independent verification is still pending, as it always is with self-reported numbers, but the sheer scale alone makes this the biggest open-source release to date.

Open Weights as a Geopolitical Flex

This isn't just a model drop — it's a statement. While US labs increasingly gate their best models behind APIs and enterprise contracts, Moonshot is handing out the keys to a 2.8-trillion-parameter engine for free download, which is either generous or a very calculated way to win developer mindshare while export controls try to slow everyone else down.

Moonshot is reportedly raising a new round that would value the company at $31.5 billion, up from $20 billion just two months ago — proof that "give away your frontier model" is, improbably, a great fundraising pitch right now. It's also a reminder that the open-weight race isn't a consolation bracket anymore; it's where some of the most serious competition is happening.

Closed labs love to talk about safety and moats. Moonshot just shipped a model bigger than most data centers and handed out the blueprints — the moat argument gets harder to make with every trillion parameters someone gives away for free.

Source: South China Morning Post