Kimi K2
Open weightsby Moonshot AI·China·Released
Open-weights 1T-parameter MoE — agentic, long-context, the model behind Kimi Chat.
About this model
Kimi K2 (July 2025) is Moonshot AI's flagship — a 1T-parameter Mixture-of-Experts (32B activated per token) that the lab open-sourced under a modified MIT license, free for commercial use. K2 is the model behind the Kimi consumer chat app (popular in China) and is increasingly used internationally for its strong agentic + coding capabilities.
K2 was notable on release for its 65.8% SWE-bench Verified score — the highest open-weights coding agent benchmark at the time, narrowing the gap to Claude Sonnet 4 considerably. The Moonshot team has emphasised 'agentic capability' as the primary design goal, which shows in the model's tool-use behaviour.
aigpt itself uses Kimi (via OpenRouter) for the news-rewrite and tool-extraction pipelines.
Strengths
- •Top open-weights SWE-bench Verified score at launch (65.8%)
- •1T-parameter MoE with permissive licensing
- •Strong agentic + tool-use behaviour by design
- •Aggressive pricing via the official Moonshot API
- •Strong Chinese + English bilingual capability
Limitations
- •Higher latency from US/EU regions than Western providers
- •Tool-call format is Moonshot-specific, not MCP
- •Less mature compliance posture for Western enterprise
- •128K context — smaller than the original Kimi Chat that made the lab famous for long context
When to use it
- →Coding agents needing open weights
- →Tool-use workflows where R1 / V3 fall short
- →Chinese-market deployments needing top-tier open quality
- →Self-hosted alternatives to Claude for agentic workloads
Architecture & training
1T total parameters, 32B activated. The Moonshot K2 technical paper describes a custom Mixture-of-Experts variant with optimisations for agentic post-training. Trained primarily on a Chinese-English mix with substantial code and tool-use training data. The model is served from APAC infrastructure by default, with international deployments via partner providers.
Benchmarks
| Benchmark | Score | Bar |
|---|---|---|
| MATH | 73.5 | |
| MMLU | 89.5 | |
| HumanEval | 85.7 | |
| SWE-bench Verified | 65.8 |