Kimi K2.6
used in
0
stacks
Moonshot AI's flagship 1T-parameter open-weight MoE model with native multimodal input and interleaved-thinking tool use.
Kimi K2.6 is released under a Modified MIT license covering both code and weights. It is a mixture-of-experts model with 1T total parameters and 32B activated (384 experts, 8 selected per token), a 256K-token context window, and a 400M-parameter MoonViT vision encoder for image and video input. The model card highlights interleaved thinking with multi-step tool calls and improvements on end-to-end coding tasks; it is served via an OpenAI/Anthropic-compatible API on platform.moonshot.ai and deployable with vLLM, SGLang, or KTransformers.
Strengths
- Interleaved thinking with multi-step tool calling for agent workflows
- 256K-token context window
- Multimodal (image and video) input via built-in vision encoder
- OpenAI/Anthropic-compatible first-party API eases drop-in adoption
Tradeoffs
- Modified MIT license (not plain MIT/Apache) requires review for some uses
- Video chat is experimental and only available through the official API
- 1T-parameter scale limits self-hosting to large deployments