Moonshot AI has introduced Kimi K3.1 Pro, an updated model featuring 2.8 trillion parameters with 104 billion active ones. The model achieved 55 points on the Artificial Analysis Intelligence Index, securing the fourth position and outperforming all OpenAI models. A faster version, Kimi K3.1 Pro Highspeed, is also available, with both models accessible via Kimi platforms and API.
The model demonstrates significant improvements in agentic tasks, scoring 1846 on the GDPval-AA benchmark and 65.7% on Terminal-Bench 4.0. Hallucination rates in the Kimi Hallu Bench Hard test dropped to 8.3% from 23.4% in the previous version. Additionally, the model supports direct integration with game engines like Unity and Unreal Engine, and includes a new 3D Make feature for 3D model manipulation.
Operational costs for the model have decreased by 33%. The price is now $2 per 1 million input tokens and $10 per 1 million output tokens, with cached data processing costing $0.20 per 1 million tokens. Moonshot AI plans to release the model weights on November 4, positioning it as a powerful open-weights system.
During development, the model showed advanced capabilities in identifying cybersecurity vulnerabilities, scoring 90.5% in the xdev-bench-hard test. To prevent misuse, Moonshot AI removed cybersecurity-related data from the version intended for public release, which reduced its performance in that specific benchmark to 81.0%.
