Kimi released its latest open‑source model, K2.6, through Kimi.com, an app, API, and Kimi Code. The update delivers advanced long‑horizon coding, agent swarm functions, and tool integration.

K2.6 climbs past its predecessor on internal benchmarks, scoring a 12% lift in code accuracy and an 18% jump in long‑context stability. It also achieved a 96.60% success rate in tool invocation.

A highlight came when K2.6 locally deployed a Qwen3.5‑0.8B model on a Mac, boosting inference throughput from ~15 to ~193 tokens/sec, outpacing LM Studio by 20%. The model also overhauled an 8‑year‑old financial engine, raising throughput from 0.43 to 1.24 MT/s and doubling performance.

Across front‑end, dev‑ops, and full‑stack tasks, K2.6 demonstrates SOTA‑level reliability and instruction following, making it a solid foundation for autonomous agent pipelines and complex engineering projects.