Kimi

Kimi is the AI product and model family from Moonshot AI, a Beijing-based AI company founded in 2023. Kimi encompasses a consumer chat assistant (kimi.ai), a developer API platform (platform.moonshot.ai), and a growing portfolio of open-weight model releases that have made Moonshot AI one of the most prominent open-source LLM providers in China and internationally.

Model families

Kimi K1.5 (January 2025) was Moonshot AI's first public reasoning model, designed to match OpenAI o1 on mathematics, coding, and multimodal reasoning benchmarks. K1.5 supports a 128K context window and uses long chain-of-thought (CoT) processing. It is a proprietary model accessible via the Kimi API.

Kimi-VL (April 2025) is an open-weight Mixture-of-Experts vision-language model with 16B total parameters (2.8B active) and a native-resolution visual encoder called MoonViT. It supports a 128K token context window, handles text, images, and documents, and competes with GPT-4o-mini and Qwen2.5-VL on multimodal reasoning benchmarks. Kimi-VL is released under the MIT license and available on Hugging Face (moonshotai/Kimi-VL, ~1.2k GitHub stars). A reasoning variant, Kimi-VL-Thinking, was released in June 2025.

Kimi K2 (July 2025) is a 1 trillion parameter MoE model with 32B active parameters, trained on 15.5 trillion tokens using the Muon optimizer. It was open-sourced under a modified MIT license and released on Hugging Face (moonshotai/Kimi-K2-Instruct, 10.7k+ GitHub stars). K2 excels at tool use, agentic reasoning, and coding — achieving top results on SWE-bench Verified. An updated release (Kimi-K2-Instruct-0905, September 2025) extended the context window from 128K to 256K tokens and improved coding performance further.

Kimi K2.5 (January 2026) added native multimodal vision to the K2 base via MoonViT (a 400M-parameter vision encoder), enabling the model to process images and video. K2.5 introduced Agent Swarm capabilities — up to 100 sub-agents coordinating across 1,500 steps — and was trained on ~15 trillion mixed visual and text tokens. Available on Hugging Face (moonshotai/Kimi-K2.5) and via the Kimi API.

Kimi K2.6 (April 2026) is the latest release in the K2 family, maintaining the 1T-parameter MoE architecture (32B active, 384 experts, 8 selected per token) with a focus on long-horizon coding and advanced agent orchestration. K2.6 scales Agent Swarm to 300 sub-agents and 4,000 coordinated steps, supports full-stack application generation, and ships as the default backend for Kimi Code CLI. Available open-source on Hugging Face and GitHub.

Kimi Linear (October 2025) is a smaller 48B parameter MoE model (3B active) that replaces standard attention with Kimi Delta Attention (KDA), a more efficient attention mechanism suited for longer sequences with reduced compute.

Kimi Dev (June 2025) is a 72B coding-focused model based on Qwen2.5-72B, open-sourced under MIT for software engineering tasks.

Access methods

Open-weight (self-hostable): Kimi-VL, Kimi K2, Kimi K2.5, Kimi K2.6, Kimi Dev, and Kimi Linear are all publicly released on Hugging Face under the MoonshotAI organization. Developers can download weights and run them on their own infrastructure. License varies by model — primarily modified MIT.

Moonshot AI API: The developer platform at platform.moonshot.ai (formerly platform.moonshot.cn) provides OpenAI-compatible REST API access at endpoint api.moonshot.ai/v1. The proprietary moonshot-v1 model series is available in three context-length variants: moonshot-v1-8k, moonshot-v1-32k, and moonshot-v1-128k. Kimi K2 and K2.5 are also available via the API. Pricing is per token (approximately $0.60/$2.50 per million input/output tokens for K2; kimi-latest selects pricing tier automatically by context length). Automatic context caching reduces costs by ~75% on repeated content.

Consumer chat (kimi.ai): The Kimi chat assistant offers a free tier with access to the latest Kimi models, supporting long documents, images, and web search. A paid plan unlocks higher usage limits and priority access to frontier models.

Kimi Code CLI: An open-source terminal-first coding agent (Apache 2.0, 6,400+ GitHub stars) that uses K2.5/K2.6 as its backend. Supports agentic multi-step coding tasks, file editing, and repo-level changes.