Kimi K2.5: Moonshot AI Enters the Multimodal Frontier

Moonshot AI has released Kimi K2.5, a 1-trillion-parameter Mixture-of-Experts model with 32 billion active parameters. The model is natively multimodal, processing both text and images within a single architecture rather than bolting on a separate vision encoder.

The benchmark numbers are strong. Kimi K2.5 scores 76.8 on SWE-bench Verified, 96.1 on AIME 2025, 98 on MATH-500, and 99 on HumanEval. The SWE-bench result is particularly notable, as it places the model among the top three globally for real-world software engineering tasks.

On the pricing side, Kimi K2.5 is available through OpenRouter at $0.45 per million input tokens and $2.20 per million output tokens. The 256K context window supports long document analysis and complex coding tasks.

What makes this release significant is the company behind it. Moonshot AI, based in Beijing, has rapidly moved from releasing capable but niche models to competing directly with Anthropic, OpenAI, and Google at the frontier. The company's earlier Kimi K2 Thinking model demonstrated strong reasoning capabilities, and K2.5 builds on that foundation with improved multimodal understanding.