Moonshot AI releases new AI model and open-source coding agent

Moonshot AI released Kimi K2.5, a native multimodal large language model, and Kimi Code, an open-source coding agent, on January 27, 2026. Kimi K2.5 is a mixture-of-experts (MoE) model with 1 trillion total parameters and 32 billion active parameters. It has a 256K-token context window and was trained on about 15 trillion mixed visual and text tokens.

The model supports agent swarms that self-direct up to 100 sub-agents. Swarms can make up to 1,500 parallel tool calls and reduce execution time by up to 4.5× for complex workflows. Kimi K2.5 generates code from images and videos and supports visual debugging.

Benchmarks show 76.8% on SWE-Bench Verified, 50.2% on Humanity's Last Exam (HLE) with tools, 73.0% on SWE-Bench Multilingual, and 86.6% on VideoMMMU. These results place it alongside GPT-5.2, Claude 4.5 Opus, and Gemini 3 Pro on key coding and vision measures.

Kimi Code runs in terminals and integrates with IDEs such as VS Code, Cursor, and Zed. It accepts multimodal inputs for end-to-end development workflows. The model is available on Hugging Face under a modified MIT license. API pricing is $0.60 per million input tokens ($0.10 cached) and $3 per million output tokens.

Why It Matters

• Teams can build visual, agentic workflows with an open-source model that matches closed-source performance on coding and vision benchmarks, enabling in-house multimodal solutions.
• Kimi Code converts images and videos into interactive code and supports visual debugging via terminal and IDE integrations (VS Code, Cursor, Zed), speeding frontend prototyping and fixes.
• Agent swarms let teams parallelize tasks and tool calls—up to 100 sub-agents and 1,500 parallel calls—cutting runtime by up to 4.5× for research, batch processing, and content generation.
• Permissive licensing and cost-effective API pricing lower barriers to scale multimodal AI in production environments.