Moonshot's Kimi K3 is now the world's largest open AI model

Moonshot AI has released Kimi K3, a 2.8-trillion-parameter model the Beijing-based startup calls the world's largest open-weight AI system. The launch lands as Chinese developers close in on their American counterparts faster than many Western analysts expected.

Kimi K3 carries a 1-million-token context window, letting it hold far more information in a single prompt than earlier Kimi releases. Moonshot says the model targets advanced reasoning, long-horizon coding, and knowledge work, and positions it as the first model to approach the 3-trillion-parameter mark.

The company has made K3 available through its work and API, though the full open-weight files, the part that lets developers download and modify the model directly, aren't due until July 27.

According to the firm, K3 "outperformed" Anthropic's Fable 5, outperformed Anthropic's Opus 4.8, and outperformed OpenAI's GPT-5.6 Sol and GPT-5.5 on GPU kernel optimisation, an indicator of hardware efficiency.

Independent assessors have given various but impressive feedback, Arena.ai put K3 in the first place in terms of creating web interfaces, Vals AI made K3 second after Fable 5, and Artificial Analysis found the performance of K3 similar to that of GPT-5.5 and Opus 4.8 in complex multistep tasks.

These numbers are mostly taken from the benchmarking conducted by Moonshot itself.

The announcement came with some shock to the domestic competition, whose stocks in Zhipu and MiniMax reportedly tumbled in Hong Kong exchanges on the very same day. And that trend continues since Zai's GLM-5.2 recently placed close to the top of US closed-source models, refuting any expectations that laboratories in China were at least six months behind those developments. Prior to the K3 announcement, Meituan’s LongCat-2.0 and DeepSeek’s V4-Pro were leading in China with 1.6 trillion parameters each.

Alibaba- and Tencent-backed Moonshot was reportedly raising $2 billion at a $30 billion valuation before listing in Hong Kong. The plans include developing their own 2.7-trillion parameter model by 2026.