Kimi K2.6 open model

Kimi K2.6 is an open model from Moonshot AI built for coding, interface generation, and tool-using agents, with weights posted on Hugging Face and support in Moonshot’s API docs. (huggingface.co) (platform.kimi.ai) The model card describes K2.6 as a “native multimodal agentic model,” which means it can take text and images, call tools, and work through multi-step software tasks instead of only answering chat prompts. Moonshot says it targets long-horizon coding, coding-driven design, and autonomous execution. (huggingface.co) (platform.kimi.ai) Under the hood, K2.6 uses a mixture-of-experts design: 1 trillion total parameters, 32 billion active parameters per token, 384 experts, and a 256,000-token context window, according to Moonshot’s model card. The same model card lists a 400 million-parameter vision encoder for image inputs. (huggingface.co) Moonshot’s quickstart says K2.6 supports text, image, and video input, plus “thinking” and non-thinking modes for different latency and reasoning tradeoffs. The company also says its API is compatible with the OpenAI software development kit format, which lowers the work needed to swap the model into existing apps. (platform.kimi.ai) That matters for developers building side projects because open-weight models can run outside a closed chatbot product, including on self-hosted stacks and private workflows. The vLLM project added a Kimi-K2.6 recipe updated on April 20, 2026, with support for tool calling, reasoning parsing, and a text-only mode that skips the vision encoder to save memory. (github.com) Moonshot is positioning K2.6 as a step beyond its earlier K2 line, which already used the same 1 trillion-parameter, 32 billion-active-parameter mixture-of-experts setup but shipped with a 128,000-token context window in the public GitHub repository. K2.6 doubles that listed context length to 256,000 tokens in Moonshot’s current documentation. (github.com) (huggingface.co) The company’s own benchmarks center on software and agent tasks rather than general chat. In the K2.6 model card, Moonshot reports 58.6 on SWE-Bench Pro, 80.2 on SWE-Bench Verified, 66.7 on Terminal-Bench 2.0, and 73.1 on OSWorld-Verified, alongside search-heavy scores such as 92.5 F1 on DeepSearchQA. (huggingface.co) Those numbers come from Moonshot, not an independent lab, and benchmark tables in model cards often depend on tool access, prompting choices, and evaluation settings. Still, the documents and deployment recipes point to the same use case: a model meant to write code, generate front ends, and act through tools in longer product-building loops. (huggingface.co) (github.com) For developers who want a model they can wire into local prototypes or private assistants, Kimi K2.6 is now showing up in the places that matter: public weights, public docs, and inference recipes that make it easier to run. (huggingface.co) (platform.kimi.ai) (github.com)