Moonshot AI launched Kimi K3.1 Pro on 23 October 2026, the first model in its K3.1 series, and published the full weights on Hugging Face on 4 November as moonshotai/Kimi-K3.1-Pro. A faster sibling, Kimi K3.1 Pro Highspeed, runs two to three times faster according to Moonshot.
What it is
| Kimi K3.1 Pro | |
|---|---|
| Architecture | Mixture of experts, same design as Kimi K3 |
| Parameters | About 2.8 trillion total, about 104 billion active per token |
| Context | Up to 1M tokens |
| Weights | Hugging Face, MXFP4, about 1.56 TB |
| Licence | Modified MIT (conditions for very large services) |
Where it ranks
On Arena's Agent leaderboard, Kimi K3.1 Pro ranks first among open-weight models and 13th overall. That makes it the strongest open model for agent work on that leaderboard at the time of writing. Moonshot's own benchmark figures are higher, but haven't been reproduced independently yet.
The licence
The Kimi K3.1 Pro License is a modified MIT licence. Use, modification and commercial use are allowed. Products and services above a very large scale (reported as over 100 million monthly users) carry extra conditions, so read the licence on the model page if you are building something big.
Running it yourself
- Hardware: at about 1.56 TB even in a 4-bit format, it needs a multi-GPU server with well over a terabyte of GPU memory. This is a model for data centres and inference providers, not a workstation.
- Safety change: Moonshot says it removed cyber-attack related data from training before the open release.
- Hosted: Moonshot's own API serves it; launch coverage lists $2 input and $10 output per million tokens.