Moonshot open-sources Kimi K2.6: 300 sub-agents and 13 hours of autonomous coding

Moonshot AI has released Kimi K2.6 in open access, an update to its flagship model for programming and agentic scenarios. The model already runs in Kimi.ai, the mobile app, via the API and in Kimi Code. The authors' main argument is that K2.6 can work on an engineering task for hours without human intervention. In one demo, over 12 hours and more than 4000 tool calls, it raised the speed of a homegrown inference engine from ~15 to ~193 tokens per second. The agent swarm, meanwhile, grew from 100 to 300 sub-agents.

Long-horizon coding

The first case is an exotic one. K2.6 downloaded Qwen3.5-0.8B onto a Mac and implemented inference in Zig, a language with a tiny share of training data. Over 14 iterations, throughput grew from ~15 to ~193 tokens/s, roughly 20% faster than LM Studio. Moonshot calls this out-of-distribution generalization, and the example is indeed telling, provided you take it at face value.

The second case is closer to production. The model took on exchange-core, an 8-year-old open-source order-matching engine. Over 13 hours it tried 12 optimization strategies, made more than 1000 tool calls and changed over 4000 lines of code. To find bottlenecks, it read CPU and allocation flame graphs, then reworked the thread topology from 4ME+2RE to 2ME+1RE. According to the authors, the results are:

• medium throughput rose 185%, from 0.43 to 1.24 MT/s;
• performance throughput rose 133%, from 1.23 to 2.86 MT/s;
• the engine was already running close to its limits before this.

Partner feedback paints a more down-to-earth picture. In its own tests, CodeBuddy sees +12% in code generation accuracy, +18% in long-context stability and 96.60% successful tool calls. A platform that serves the model through AI Gateway records a gain of more than 50% on its own Next.js benchmark. Another partner reports +15% on its tests. All of these figures come from the partner companies' internal test sets. They cannot be compared with one another, and there is no independent verification yet.

A swarm of 300 agents and Claw Groups

The agentic part of the release is built on Agent Swarm. The system splits a task into heterogeneous subtasks and hands them to agents that it spawns itself. In K2.6, the swarm scales to 300 sub-agents and 4000 coordinated steps, versus 100 and 1500 in K2.5. The claimed result of a single autonomous run is documents, websites, presentations and spreadsheets together. The model can turn any good PDF, spreadsheet or Word file into a Skill, preserving the document's structure and style for future tasks.

For always-on agents such as OpenClaw and Hermes, Moonshot cites an experiment: an RL infrastructure team ran an agent on K2.6 for 5 days, handling monitoring, incident response and operations from alert to resolution. The new feature, Claw Groups, ships as a research preview. It lets you plug agents from any device and running any model into a shared working loop, each with its own skills and memory. K2.6 acts as the coordinator: it distributes tasks according to skill profiles, spots failures and reassigns work. Inside the company, a marketing team of agents works this way, including Demo Makers, Benchmark Makers, Social Media Agents and Video Makers.

How to check this in practice

Moonshot ran all comparisons against Claude Opus 4.6 (max effort), GPT-5.4 (xhigh) and Gemini 3.1 Pro (high thinking). For K2.6 it used temperature 1.0, top-p 1.0 and a context of 262 144 tokens. On the text-only subset of Humanity's Last Exam, the model scores 36.4% without tools and 55.5% with them. It makes sense to recheck the results of the internal Kimi Code Bench, Kimi Design Bench and Claw Bench on your own tasks.

If you run K2.6 through a third-party provider, check Kimi Vendor Verifier: Moonshot recommends it for choosing high-accuracy services. It advises reproducing the official figures through the official API. Teams building agentic pipelines should start with long multi-step tasks such as refactoring or profiling. That is where the model shows its biggest lead over K2.5, and the token cost of an open model makes it affordable to run such sessions for hours.