Moonshot launched Kimi K3 on July 16, and the interesting part is not just the 2.8 trillion parameters. K3 is a sparse MoE model — it activates 16 of 896 experts per token — with native vision and a 1M-token context window, aimed directly at long coding runs, research, and knowledge work. It is live in Kimi, Kimi Work, Kimi Code, and the API. Moonshot subsequently published the full weights under the Kimi K3 license.
The price is no longer “shockingly cheap Chinese model” territory: $3 / $15 per million cache-miss input/output tokens, with cached input at $0.30. That is closer to Sonnet pricing than the Kimi K2-era bargain basement. But it stays flat across the full context window, and automatic prefix caching changes the math for coding agents that keep a repository or long spec stable across turns.
Moonshot's charts put K3 near Claude Fable 5 and GPT-5.6 Sol, sometimes ahead, on agentic coding and knowledge-work evaluations. Read the footnotes before treating that as a clean leaderboard win: K3, Claude, and OpenAI results were collected with different harnesses, and some competitor figures are adopted from their own release material. Moonshot is unusually clear about the practical caveats too. K3 expects its full thinking history to be preserved across turns, can get unstable if you switch it into an existing session, and has a habit of making decisions on ambiguous work unless you constrain it hard.
Practical read: K3 is an open-weight option for long agent loops, but publishing weights does not make self-hosting simple. Moonshot recommends supernode deployments with 64 or more accelerators, so most teams should evaluate the hosted API before planning infrastructure. The checkpoint also does not make Moonshot's own benchmark results independent evidence; compare K3 on your real harness and acceptance tests.