Kimi

Kimi is Moonshot AI's family of large language models and the assistant built on them. Moonshot launched Kimi in October 2023 and released Kimi K2, its first trillion-parameter model, as open weights in July 2025;¹ the K2 card lists 1 trillion total and 32 billion activated parameters in a mixture-of-experts design, under a modified MIT license.² Kimi K2.6 followed on April 20, 2026 and Kimi K3 on July 16, 2026.³ Moonshot describes K3 as a 2.8-trillion-parameter model built on Kimi Delta Attention and Attention Residuals, with native vision, a 1-million-token context window and 16 of 896 experts active, and said its full weights would be released by July 27, 2026; its launch post adds that K3's overall performance "still trails" the strongest proprietary models.⁴ K3 is published on Hugging Face under the Kimi K3 License,⁵ and the Kimi API serves K3 with a 1,048,576-token context window and K2.6 with 262,144.⁶

In SWE-bench Verified, Pro and What Replaced Them: By the Numbers, Kimi K2.5 resolves 70.8% of SWE-bench Verified tasks in the shared mini-SWE-agent v2.0.0 harness, at an average cost of $0.15 per task, in a run dated February 17, 2026.⁷ On price, K3 lists $15.00 per 1M output tokens and K2.6 lists $4.00, 17th and 11th of the 19 models in Cheapest LLM APIs per 1M Output Tokens.⁶ In Chinese Open-Weight Models: Downloads, Derivatives and Token Share, the Kimi-K3 repo logged 1,117,095 downloads in the 30 days to October 13, 2026 and has 11,707 likes, the most of any repo in that post's table; its license requires products with more than 100 million monthly active users or $20 million in monthly revenue to display "Kimi K3."⁵