Moonshot open-sources Kimi K2.6 for long-horizon coding
Moonshot AI has released Kimi K2.6 as an open-weight model aimed at long-running agentic coding, citing runs of 12 hours or more on its own demos, while most supporting evidence comes from internal tests and partners.
Moonshot AI announced that it is open-sourcing Kimi K2.6, the successor to K2.5. The company is pitching the model on coding, long-horizon execution and "agent swarm" capabilities. K2.6 is available through Kimi.com, the Kimi App, the API and Kimi Code.
Moonshot says K2.6 improves significantly on K2.5 on Kimi Code Bench, its internal benchmark of end-to-end coding tasks. It also says the gains carry across languages such as Rust, Go and Python, and across work including front-end, devops and performance optimisation. The announcement gives no scores for that benchmark.
The company described two long-running demonstrations:
- Zig inference engine: K2.6 deployed Qwen3.5-0.8B locally on a Mac and wrote and tuned inference code in Zig. Over 4,000+ tool calls, 12+ hours and 14 iterations, throughput rose from about 15 to about 193 tokens/sec. Moonshot says that is roughly 20% faster than LM Studio.
- exchange-core: In a 13-hour run, the model changed more than 4,000 lines of an eight-year-old open-source matching engine, using 1,000+ tool calls. It reworked the thread topology from 4ME+2RE to 2ME+1RE. Medium throughput rose from 0.43 to 1.24 MT/s and peak throughput from 1.23 to 2.86 MT/s.
Beta partners also gave feedback. Tencent's CodeBuddy reported 12% higher code-generation accuracy than K2.5, 18% better long-context stability and a 96.60% tool-call success rate. Augment Code, Ollama and OpenCode praised its reliability over multi-step sessions.
The post seen here does not give model size, license terms or pricing. None of the figures above has been independently verified.