Moonshot releases open-weight Kimi K2.6 for agentic workflows

Moonshot AI has released Kimi K2.6, a new open-weight model aimed at agentic workflows, and Kilo Code says it is already live in its tools. The company says the model is designed for long-running coding tasks, multi-agent systems, and other work that depends on sustained context across large codebases.

Kilo Code co-founder and CEO Scott Breitenother said K2.6 is now available in Kilo Code and KiloClaw, the company’s hosted OpenClaw service. In the announcement, he said early testing showed strong performance on long-context tasks across large projects, along with lower cost than frontier closed models.

Moonshot AI’s release follows Kimi K2.5, which Kilo users had praised for reasoning through complex codebases, suggesting refactoring strategies, and maintaining context in large projects. According to the post, K2.6 continues that direction and is meant to stay competitive with top-tier proprietary models.

The benchmark numbers cited in the release are aimed squarely at software engineering and agentic work. K2.6 scored 80.2% on SWE-Bench Verified and 58.6% on SWE-Bench Pro, along with 92.5% F1 on DeepSearchQA and 66.7% on Terminal-Bench 2.0.

The Kilo Code post also highlights the model’s endurance in extended runs. In one 13-hour execution period, Kimi K2.6 reportedly iterated through 12 optimization strategies, made more than 1,000 tool calls, and modified more than 4,000 lines of code, leading to a 185% increase in median throughput from 0.43 to 1.24 MT/s.

That kind of sustained execution is the main selling point for teams running always-on agents. Breitenbreother said K2.6 is well suited for deep codebase refactoring, non-obvious bug hunting, and autonomous workflows that need to keep going without constant human intervention.

The model also expands the scale of multi-agent setups. According to the post, K2.6 can horizontally scale to 300 sub-agents across 4,000 coordinated steps, up from K2.5’s limit of 100 sub-agents and 1,500 steps.

Kilo Code says K2.6 can also turn files such as PDFs, spreadsheets, slides, and Word documents into agent skills, broadening its use beyond code. Michael Chiang, co-founder of Ollama, said in the announcement that the model raises the bar for open-source models and works well for agentic tools like OpenClaw and Hermes.

The company also warns that the model can be very creative, so users should give it clear instructions. Kimi K2.6 is available now through the Kilo Gateway and can be used in the Kilo CLI, the company’s VS Code and JetBrains extensions, Hermes, KiloClaw, and other Kilo-connected tools.