Moonshot’s Kimi K3 is the world’s largest open AI model

TL;DR:

• Beijing-based Moonshot released Kimi K3, a 2.8 trillion-parameter open-weight model with a 1 million-token context window, claiming performance approaching Anthropic’s Fable.
• Third-party evaluations broadly support the claim: Vals AI ranked it second overall behind Fable 5, and Arena.ai placed it first for web interface-building.
• The release accelerates a pattern of Chinese open models closing on US frontier systems, weeks after the US government temporarily withdrew Anthropic’s most advanced models.

Chinese AI startup Moonshot has unveiled Kimi K3, describing it as the first open-weight model to approach three trillion parameters and the largest open AI system released anywhere. The company says it is built for advanced reasoning, long-horizon coding and knowledge work, with two architectural changes that improve computing efficiency and let it complete extended coding tasks with minimal supervision.

The benchmark claims are notable for their independent backing. Vals AI placed Kimi K3 second overall, behind Anthropic’s Fable 5 but ahead of OpenAI’s GPT-5.6 Sol; Arena.ai ranked it first for building web interfaces; and Artificial Analysis judged it comparable to GPT-5.5 and Claude Opus 4.8 on complex multi-step tasks.

The timing carries weight. The launch comes a month after the US government abruptly withdrew Anthropic’s Fable and Mythos models over security concerns — restrictions since lifted — and follows Z.ai’s GLM-5.2, which undermined the Western consensus that Chinese models trailed by at least six months. MiniMax plans its own 2.7 trillion-parameter model as soon as the third quarter. Markets read the release as a competitive shock: shares in rivals Zhipu and MiniMax fell 21.9% and 13.8% in Hong Kong by midday.

The open-weight trade-off

For UK businesses, ever-larger open-weight models cut the cost of running capable AI on their own infrastructure — a point underlined by Thinking Machines’ giant open model Inkling this week. But openness cuts both ways: a UK researcher has just demonstrated an open-weight model can be backdoored for under £75, and provenance checks lag far behind those for conventional software.

Looking forward

Moonshot, backed by Alibaba and Tencent, was reported last month to be seeking $2 billion at a $30 billion valuation ahead of a potential Hong Kong listing. With trillion-parameter releases now arriving quarterly from Chinese labs, the question is whether US frontier developers keep their shrinking lead — and at what price advantage.