What it is. On 16 July, Moonshot AI of Beijing released Kimi K3: 2.8 trillion parameters — the largest open-weight model yet by parameter count — natively multimodal, with a one-million-token context window, and aimed at long-range programming, knowledge work and reasoning. One precision matters. The weights were not public at launch; Moonshot has promised them by 27 July, and until they land the model is open-weight in commitment rather than in fact.
How good, independently. Artificial Analysis, the field’s most-watched independent evaluator, places K3 fourth of 189 models on its composite Intelligence Index — comparable to Claude Opus 4.8 and GPT-5.5, behind only Claude Fable 5 and GPT-5.6 Sol. Moonshot itself concedes the same ordering.
Good in what sense. Agentic, long-horizon work is the franchise. On GDPval, a benchmark of real-world occupational tasks, K3 reaches an Elo rating of 1,668 against its predecessor’s 1,190 — passing GLM-5.2, GPT-5.5 and Opus 4.8, trailing only Fable 5. It places second on the evaluator’s long-horizon knowledge-work benchmark, first on its automation benchmark, and topped a global front-end programming leaderboard within a day of release.
The caveats. At $3 and $15 per million input and output tokens, it is the dearest model any Chinese laboratory has shipped — Western mid-range territory, working out at roughly $0.94 per completed task: about half the cost of Opus 4.8, but above its open-weight peers. It is hungry, generating around twice the median output tokens of comparable models, at 62 tokens per second — below the field. Its hallucination rate rose against its predecessor; its launch table mixes agent harnesses, which flatters comparison; and self-hosting is a serious undertaking, with a recommended deployment of some 64 accelerators.
Verdict. The strongest independently verified result yet from a Chinese laboratory and, once the weights ship, presumptively the leading open model — best-in-class for agentic work and coding at mid-tier cost, with the overall frontier still, narrowly, American. The market read it that way: shares of domestic rivals Zhipu and MiniMax fell 28.4 and 15.6 per cent on the day. And it retires a line in Under the Hood, which named K2.6 the strongest open-weight model on the composite indices — the first entry in this register doing precisely what the register is for.