the moonshot team officially released and open-sourced its latest model, kimi k2.6, yesterday, delivering comprehensive upgrades in code writing, long-term task execution, and agent cluster capabilities. starting today, all users can access the model through the official website, the latest version of the kimi app, the kimi api, and the kimi code programming assistant.
according to official disclosures, kimi k2.6 has achieved industry-leading results in tests such as the phd-level “ultimate human exam,” swe-bench pro, and deepsearchqa, performing on par with or better than closed-source models like gpt-5.4 and claude opus 4.6. as moonshot’s most powerful code model to date, k2.6 can continuously generate code for up to 13 hours, writing or modifying over 4,000 lines of code and completing complex system development. in internal benchmark tests, its performance has improved by approximately 20% compared with the previous generation.
real-world test cases show that the model successfully deployed the qwen3.5-0.8b model locally on a mac, achieving a 20% faster inference speed than lm studio after more than 4,000 tool calls and 12 hours of operation. in another case, k2.6 autonomously completed a deep refactoring of an open-source financial engine, increasing median throughput by 185%.
k2.6’s agent cluster capabilities have also been upgraded, allowing it to schedule up to 300 sub-agents to execute 4,000 collaborative steps in parallel, enabling end-to-end delivery of multiple outputs—from documents to web pages, ppts, and spreadsheets. in addition, the kimi agent mode now supports skill creation and invocation, with hundreds of officially recommended skills built-in. moonshot has also launched a small-scale beta test of the “claw group” feature, enabling multiple agents to work collaboratively with humans. k2.6 is now available to all free and paid users, and the api is simultaneously offering a limited-time promotional credit of up to 30%.
