Kimi K2

Kimi K2 is a multimodal AI model by Moonshot AI for developers, researchers, and AI enthusiasts who need long-context reasoning, coding help, and agentic task execution via web chat, API, or local deployment.

Key capabilities include:

• Native vision and language support for visual analysis and cross-modal reasoning, including visual coding that can generate code (for example, React/HTML) from UI mockups, images, or video-based workflows
• Agentic modes for autonomous work, including an Agent mode designed for extended workflows with 200 to 300 tool calls, and Agent Swarm coordination for up to 100 parallel agents
• 256K context window for long-document analysis and extended sessions with fewer context resets
• Local deployment options with community tooling support, including recommended inference engines such as vLLM, SGLang, KTransformers, and TensorRT-LLM

Pricing: Paid plans include Starter ($9/month + 10M credits monthly), and Ultra ($49/month + 70M tokens monthly).

Notable for its Mixture-of-Experts design (1T total parameters, 32B active per request), Multi-Head Latent Attention to make 256K context more efficient on standard GPUs, and open-source availability (Modified MIT License) with Base and Instruct variants for customization or chat and agent use.