kimi capabilities

Kimi's Agent mode lets a single model autonomously complete multi-step tasks — web browsing, content creation (slides, websites, docs, sheets), and code execution. Agent Swarm scales this by spinning up to 300 domain-specialized sub-agents executing up to 4,000 coordinated steps in parallel, dramatically reducing wall-clock time for large tasks.

Kimi Deep Research is an autonomous research agent that plans keywords, searches hundreds of web pages, and produces long-form reports (10,000+ words) with traceable citations and an interactive HTML visual output. Basic web search is also available inline in standard chat.

Users can upload PDFs, Word documents, spreadsheets, presentations, images, and video files into Kimi chat for analysis, summarization, and extraction. The model reads the file contents within the context window and can act on them alongside text instructions.

Kimi processes inputs up to 256K tokens in a single conversation, enabling users to submit entire codebases, lengthy legal documents, or book-length texts in one pass without truncation. Context is maintained losslessly across the full window.

Kimi is available as a free iOS and Android app that mirrors the web product's feature set, including chat, file upload, agent mode, and voice interaction. The same membership tier applies across web and mobile.

Moonshot AI releases Kimi model weights publicly under the Modified MIT License, making them freely downloadable and self-hostable. A commercial API at platform.moonshot.ai / api.moonshot.ai gives developers pay-per-token access to the latest models with OpenAI-compatible endpoints.

Kimi offers a free tier and four paid subscription tiers named after musical tempos. The free Adagio plan gives access to K2.6 chat with file upload and a daily message allowance. Paid plans (Moderato through Vivace) unlock higher agent-use quotas, Agent Swarm, Deep Research, Kimi Code credits, and concurrent task slots.

Kimi offers a dedicated Thinking mode that extends generation with explicit step-by-step chain-of-thought reasoning before delivering a final answer. It is designed for problems requiring deep analysis, complex mathematics, and multi-step logic. A faster Instant mode (no extended thinking) is also available.

Kimi K2.5 and K2.6 are natively multimodal models that understand images and (experimentally) video, enabling UI-to-code workflows, visual question answering, design-to-production conversion, and screenshot analysis. The MoonViT vision encoder processes images up to 4K resolution.