Kimi K2

Kimi K2 is the model on which Moonshot AI's entire later line stands — the catalogue already holds its descendants K2.6 and K3, and foreign products such as Sakana's Namazu are built on this family. It was published with open weights on 11 July 2025 in two forms: Kimi-K2-Base for further training, and Kimi-K2-Instruct for direct use. The architecture is a mixture of experts with one trillion total parameters and 32 billion activated per token: 61 layers, 384 experts with eight selected per token plus one shared, 64 attention heads, MLA attention, a 160K vocabulary and a 128K context window. The number the maker itself treats as the achievement is not the size but the stability: pre-training a trillion-parameter model on 15.5 trillion tokens with, in its words, zero training instability, using its own MuonClip optimiser at a scale nobody had used the Muon method at before. The second deliberate decision is what the model does not have. Moonshot describes Instruct as a reflex-grade model without long thinking — it answers immediately rather than reasoning step by step, and the whole design targets tool use and autonomous work rather than contemplation. The separate thinking branch of the family arrived only four months later, as Kimi K2 Thinking. In the maker's own comparison K2 leads the open models of its day and stays behind the closed flagships: SWE-bench Verified in agentic mode 65.8 on a single attempt and 71.6 across several, against 38.8 for DeepSeek-V3-0324 and 72.7 for Claude Sonnet 4; LiveCodeBench v6 53.7, the best result in that table. The licence is a modified MIT: free use and modification, with an attribution requirement above a commercial threshold. A later refresh, Kimi-K2-Instruct-0905, extended the served window to 262,144 tokens; this profile describes the July release, and the catalogue keeps older versions rather than overwriting them.

▸Videos

No videos yet.

▸Reviews