Technical capabilities & architecture

Built by Moonshot AI for autonomous, long-running agents

Kimi is Moonshot AI's flagship language model family. Built on a Mixture-of-Experts architecture with up to 1 trillion parameters, Kimi handles 262K token context windows and is purpose-built for agentic coding sessions that span hours and thousands of tool calls. Open weights are published on Hugging Face, making Kimi a top choice for self-hosted enterprise AI and research.

Key features

Massive context — 262K-token window for deep document and codebase analysis.
Agentic by design — multi-hour coding sessions with up to 300 parallel sub-agents.
Open weights — full models on Hugging Face for self-hosted deployments.
Top-tier benchmarks — 80%+ on SWE-Bench Verified and 93% on MathVision.

Featured models

Kimi AI · AI Models

Flagship

Kimi K2.6

1T parameter MoE built for 12-hour autonomous coding sessions, 80% on SWE-Bench Verified.

Context window 262K Released Apr 2026

Stable

Kimi K2.5

Previous-generation Kimi K2 with strong long-context coding and reasoning.

Context window 256K Released Jan 2026

Compare

Kimi AI vs Chatgpt

Kimi AI vs Chatgpt

Compare with Chatgpt.

Kimi AI vs Claude

Kimi AI vs Claude

Compare with Claude.

Kimi AI vs Deepseek

Kimi AI vs Deepseek

Compare with Deepseek.

Try Kimi AI

Kimi is Moonshot AI's flagship language model family. Built on a Mixture-of-Experts architecture with up to 1 trillion parameters, Kimi handles 262K token context windows and is purpose-built for agentic coding sessions that span hours and thousands of tool calls. Open weights are published on Hugging Face, making Kimi a top choice for self-hosted enterprise AI and research.

Open Chat