Chinese artificial intelligence lab Moonshot has released Kimi K3.1 Pro, an upgraded version of its open-weight model, claiming it comes close to matching Claude Opus 5.5 in a coding benchmark.
The company announced the release on Friday, making the model available through Kimi’s applications and application programming interface.
Moonshot plans to publish the model’s full weights by 4 November.
K3.1 Pro builds on Kimi K3, a model with nearly three trillion parameters, with additional training to improve its capabilities.
The company has also introduced a Highspeed version that it says runs two to three times faster.
Moonshot said it had removed hacking-related training data to reduce the risk of misuse when the model’s weights become publicly available.
On Terminal-Bench 4.0, a benchmark for AI coding agents, Kimi K3.1 Pro scored 65.7 per cent, according to Moonshot. The result puts it ahead of OpenAI’s GPT-6 Astra, which scored 58.2 per cent, and just behind Anthropic’s Claude Opus 5.5 at 66.4 per cent.
The performance figures were provided by Moonshot and have not been independently verified.
The company also claims the upgraded model is more accurate, with the rate of false claims falling to 8.3 per cent from 23.4 per cent for Kimi K3 in its internal stress test.
Moonshot said it removed cyberattack-related training material before releasing the model’s weights publicly, in a move aimed at reducing the risk of misuse.
The decision reflects the company’s effort to limit the cybersecurity risks associated with making powerful AI models freely available.
