Strengths
- It scored 60 on the Artificial Analysis Intelligence Index in August 2026, the highest of any open-weight model.
- A huge open model with 2.8 trillion parameters, a one-million-token context and image and video understanding.
- Moonshot reports strong agent results: 91.2 on the BrowseComp web browsing test and 84.8 on the OSWorld-Verified computer use test.
- The license conditions tied to revenue and user numbers do not apply to internal company use.
Weaknesses
- In the API it costs more than Kimi K3.1 Pro and most Chinese rivals: $3 input and $15 output per million tokens.
- Running it yourself is practically out of reach for a small company, as industry press says it needs 64 or more accelerators.
- The license is not plain MIT: large service providers need a separate agreement and very large products must display the Kimi K3 name.
- Thinking is always on.
Best for
- running a top open model on your own GPU cluster
- agents that browse the web and operate a computer
Not ideal for
- low-cost API use, as K3.1 Pro is cheaper
- small company servers
How to use it in Europe
You can use the model on Kimi.com, in the Kimi mobile app and through the international API in USD. Moonshot may store API data on servers in Singapore and does not offer EU processing. You can download the weights from Hugging Face (moonshotai/Kimi-K3) and keep all data in house on your own servers. The model is so large, though, that only an organization with data-center hardware can manage it.
Price compared with similar models
Benchmark results
| Benchmark | Score | Source |
|---|---|---|
| Artificial Analysis Intelligence Index | 60 (#1 among open-weights models) | deeplearning.ai ↗, August 2026 |
| Terminal-Bench 2.1 | 88.3 | huggingface.co ↗ (vendor-reported), July 2026 |
| GPQA Diamond | 93.5 | huggingface.co ↗ (vendor-reported), July 2026 |
| BrowseComp | 91.2 | huggingface.co ↗ (vendor-reported), July 2026 |
| OSWorld-Verified | 84.8 | huggingface.co ↗ (vendor-reported), July 2026 |
Benchmarks cover only a narrow slice of ability and vendors often report them at their best settings. Treat them as a rough guide.
More models from this vendor
- Kimi K3.1 Pro: Moonshot AI's newest model from October 2026, built for agentic coding, office documents and very long outputs.
Similar models from other vendors
- Claude Opus 5.5 (Anthropic): Anthropic's main model for demanding work, coding and long documents, and the one Anthropic recommends starting with.
- DeepSeek V4 Pro (DeepSeek): DeepSeek's largest open-weight model for demanding text tasks at a low price. It does not understand images.
- Gemini 4 Carbon (Google): Google's most capable model, announced on 28 October 2026. So far only selected cybersecurity experts have access.
- Gemini 4 Argon (Google): Google's most capable publicly available model, out in October 2026. Built for demanding analytical work where accuracy matters and made-up answers are costly.
- Muse Wonder (Meta): Meta's strongest model for long-running agents and coding, still released as a preview.
- Mistral Large 4 (Mistral AI): Mistral's largest model, still in preview. Built for demanding tasks, long documents and security analysis with data kept in the EU.