Moonshot AI Releases Kimi K2.6: Open-Source Multimodal Agentic Model Pushes Boundaries in Long-Horizon Coding and Agent Swarms

Moonshot AI has released Kimi K2.6, an updated version of its open-source flagship model. The company says the new version is focused on practical applications in coding, design, and autonomous task management. It is available now on Kimi.com, the Kimi mobile app, and through the company’s API and Kimi Code tool. The model weights are also hosted on Hugging Face under a modified MIT license.

K2.6 uses a native multimodal Mixture-of-Experts architecture. According to Moonshot AI, it has 1 trillion total parameters with 32 billion activated per token, drawn from 8 active experts plus 1 shared expert across 384 total experts. The model has 61 layers, a 256,000-token context window, and a 400-million-parameter MoonViT vision encoder. It accepts text, images, and video, and users can choose between Thinking and Instant response modes.

Coding and design

Moonshot AI says K2.6 is built for sustained programming work. The company demonstrated the model making more than 4,000 tool calls across a session that ran for over 12 hours. In one example, it reportedly downloaded and optimized inference for the Qwen3.5-0.8B model using Zig on a Mac, increasing throughput from roughly 15 to about 193 tokens per second. In another test, it analyzed and refactored an eight-year-old financial matching engine; Moonshot claims this produced throughput gains of up to 185 percent after bottleneck fixes and core code changes across thousands of lines.

For design, the model can generate front-end interfaces from simple text prompts or visual inputs. The output can include interactive elements, animations, video backgrounds, and lightweight full-stack features such as authentication and databases.

Agent capabilities

Moonshot AI says K2.6 can scale to 300 parallel sub-agents handling 4,000 coordinated steps for end-to-end workflows such as document analysis, writing, and generating websites, slides, and spreadsheets. Users can also convert uploaded files—PDFs, spreadsheets, and others—into reusable templates that keep their original style and structure.

The company also highlighted autonomous background agents, noting internal tests in which a K2.6 agent ran for five days managing monitoring and incident response. The model supports third-party agent frameworks including OpenClaw and Hermes.

Benchmark results

Moonshot reported the following scores on industry tests:

• Agentic: HLE-Full (with tools) 54.0%, BrowseComp 83.2%, Claw Eval (pass^3) 62.3%

• Coding: SWE-Bench Verified 80.2%, SWE-Bench Pro 58.6%, SWE-Multilingual 76.7%, Terminal-Bench 2.0 66.7%

The company says these results improve on the previous K2.5 model and are comparable to—or better than—those of proprietary models such as GPT-5.4 and Claude Opus 4.6 on some agentic and coding tasks, while priced lower on certain platforms like OpenRouter.

Availability

Developers can access K2.6 through an API compatible with OpenAI and Anthropic formats. Moonshot says the model is optimized for inference engines including vLLM, SGLang, and KTransformers, and is available at launch through partners such as Fireworks AI and Cloudflare Workers AI. A “Vendor Verifier” tool lets users check whether third-party deployments match the official release.

The update also introduces Claw Groups, a feature intended to let multiple specialized agents work alongside human users on shared tasks.

Moonshot AI, which operates in China’s competitive AI market, has been iterating quickly. With K2.6, the company is emphasizing open-weight, task-oriented tools rather than scale alone. Developers and enterprises can try the model now through the Kimi website or the Hugging Face repository.