Kimi

The AI assistant from Moonshot AI: the open-source 3T-class K3 model + a 300 sub-agent cluster, with million-token context + native vision.

Kimi is the flagship AI assistant of Moonshot AI. It made its name in 2023 with “200,000-character long context,” but today its two hallmarks are open source and agent clusters. The latest flagship, K3 (July 2026), is the world’s first open-source 3T-class model — officially announced with a 2.8-trillion-parameter MoE architecture, native vision support, and a 1M-token context window, matching Claude Fable 5 and GPT-5.6 Sol on multiple Coding, Agentic, and multimodal benchmarks. The Agent Swarm cluster can dispatch up to 300 sub-agents working in parallel — officially about 4.5× faster than a single agent. K3 is live simultaneously on the Kimi main site, Kimi Code, and the Kimi Platform API, with weights available for download.

Basic conversations are free. Memberships come in four tiers — ¥49 / ¥99 / ¥199 / ¥699 per month — and heavy-duty features such as Agent clusters, Kimi Code, PPT, and Deep Research share a single monthly token-based quota pool; the international version starts at $19/month. 48 hours after K3’s launch, demand exceeded expectations and GPU capacity approached its limit; on July 19 the company announced a temporary pause on new paid subscriptions (existing subscribers are unaffected). Momentum at the company level is strong: a valuation of roughly US$31.5 billion as of mid-2026, ARR surging from $100 million to $300 million in three months, with Alibaba as a major investor.

Positioning-wise, Kimi is the most closely watched Chinese open-source model family after DeepSeek — but the two take different paths: DeepSeek leans toward “pure model + extreme cost efficiency,” while Kimi has bet its focus on productized agent capabilities (clusters, the dedicated assistant Kimi Claw, 200-300 consecutive tool calls). Long Chinese document processing remains a strength compared to competitors like Doubao. Kimi Work has been in public testing since at least early June 2026: it mounts local folders, offers WebBridge browser automation and built-in Cron scheduled tasks, and ships pre-integrated with A-share/HK-stock/US-stock financial data — entering the office agent battlefield as a “desktop digital employee.”

Note: K3 is sensitive to how thinking history is passed along — switching to another model mid-task and switching back can cause inconsistent quality; it may be overly proactive about making decisions on the user’s behalf during long tasks; API pricing is now clearly higher than DeepSeek’s (the coding version runs about 6.5× the price), so its “cheap and generous” label is fading; heavy features are concentrated in the ¥99-and-above tiers, and the free tier offers limited benefits.

Model Iteration Timeline

Date	Version	Highlights
2023-10	Kimi Chat	Launched with 200,000-character long context — an instant hit
2024-03	2M-character version	Lossless context extended to 2 million characters
2025-01	k1.5	Multimodal thinking model, benchmarked against OpenAI o1
2025-07	K2	Open-source 1-trillion-parameter MoE agentic model
2025-11	K2 Thinking	Open-source thinking model, 200-300 consecutive tool calls
2026-04	K2.6	256K context + visual input + 300 sub-agent cluster
2026-07	K3	World’s first open-source 3T-class model: 2.8T MoE + 1M context + native vision, benchmarked against Fable 5 / GPT-5.6 Sol

Source: Agentopia · Full-site machine-readable catalog at /llms.txt