Kimi drops K2.6 open-source coding model
Kimi just launched Kimi K2.6, an open-source coding model built for long-horizon engineering tasks. It supports over 4,000 sequential tool calls, runs continuously for 12 hours, and orchestrates swarms of up to 300 parallel sub-agents to refactor 100+ files at once. Model weights are live on Hugging Face alongside web access on kimi.com.
Kimi K2.6 is Moonshot AI's latest open-source coding model, purpose-built for multi-step software engineering.
What it does
The headline upgrade here is long-horizon execution. K2.6 handles over 4,000 continuous tool calls across a 12-hour runtime, spanning languages like Rust, Go, and Python for frontend, devops, and performance work.
For frontend builds, it handles WebGL shaders, Three.js 3D scenes, and animations using GSAP and Framer Motion right out of the box.
Agent orchestration also got a massive bump. K2.6 scales to 300 parallel sub-agents running 4,000 steps per task—up from K2.5's 100 agents and 1,500 steps. In practice, a single prompt can coordinate refactoring across more than 100 files at once.
On benchmarks, Moonshot claims open-source state-of-the-art marks: 58.6 on SWE-Bench Pro, 76.7 on SWE-bench Multilingual, 54.0 on HLE with tools, 83.2 on BrowseComp, and 93.2 on Math Vision with Python.
Model weights and code are already live on Hugging Face. You can also run it through chat or agent mode on kimi.com, hook it up via Moonshot's API, or pair it with Kimi Code for production builds.
Why it matters
Most open-source coding models fall apart after a dozen sequential tool calls, bleeding context and looping into errors.
By pushing execution out to 4,000 steps across swarms of 300 sub-agents, K2.6 aims straight at messy, real-world refactoring jobs rather than toy snippets. If you run complex repository migrations or autonomous background devops, this is an open-weight drop worth cloning immediately.
Sources (1)
- x.com(opens in a new tab)Originalx.com