Moonshot AI has released Kimi K2.6, an updated open source model built on a Mixture of Experts (MoE) architecture: one trillion total parameters, with 32 billion active per token. The weights are available on Hugging Face under a modified MIT license.

The stated focus of this release is handling extended coding tasks, where the model needs to maintain coherence across multi-hour sessions with hundreds of tool calls and numerous iterations. Moonshot AI documents two internal examples: optimizing inference for a Qwen model on Mac in Zig, a niche programming language, and refactoring an eight-year-old open source financial engine across over 4,000 lines of code. According to the developer, both were completed without human intervention in sessions lasting more than 12 hours.

Benchmarks and comparisons

Moonshot AI’s published results show K2.6 on par with leading closed source models across several coding agent benchmarks: on SWE-Bench Verified it reaches performance comparable to Claude Opus 4.6, while on HLE with tools it ranks first in internal testing. The benchmarks are largely conducted or commissioned by Moonshot AI itself, which calls for the usual caution.

Agents and experimental features

Compared to predecessor K2.5, K2.6 expands agent coordination capabilities: from 100 to 300 specialized agents running in parallel, and from 1,500 to 4,000 coordinated steps. A research preview feature called “Claw Groups” lets you integrate heterogeneous agents, running on different devices and models, within a workflow coordinated by K2.6.

The model is accessible at Kimi.com, via API at platform.kimi.ai, and through Ollama for local use.