Overview
- Moonshot released and open-sourced Kimi K2.6, now available on the Kimi site and app, through the Kimi API, and in the Kimi Code assistant, with the official API also listed on Tencent Cloud’s TokenHub and a limited-time credit promotion.
- The company says K2.6 leads key tests that probe advanced reasoning, real-world software bug fixing, and deep retrieval, reporting top scores on Humanity’s Last Exam, SWE-Bench Pro, and DeepSearchQA versus GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro, though broad third-party validation remains limited.
- K2.6 is pitched for long-running coding work and, in a disclosed run, worked for about 12 hours, triggered over 4,000 tool calls, and used the Zig language to speed a local Qwen3.5-0.8B setup on a Mac from about 15 tokens per second to about 193 tokens per second, which the team reports is 20% faster than LM Studio.
- Agent features now include swarms that can spin up to 300 specialized workers for roughly 4,000 coordinated steps, and an internal test showed a K2.6-based operations agent running for five days straight to monitor systems, respond to alerts, and carry fixes to completion.
- Hiring moves and product trials point to a shift from pure model research to building an inference platform and developer ecosystem, with new roles focused on agent engineering and lower formal education bars to draw hands-on infrastructure talent.