Moonshot AI announced this week Kimi K3.1 Pro, the first model in the K3.1 series and the most capable of its open-weight class with approximately 3 trillion parameters. It is a post-trained version of the July Kimi K3, with significantly more post-training compute.
What the model offers
Kimi K3.1 Pro has approximately 2.8 trillion total parameters and 104 billion active per token (MoE architecture), with a 1 million token context window. It accepts text, images, and video, and always reasons before answering. The company also launched a Highspeed variant that is 2 to 3 times faster.
In independent benchmarks, the model scored 55 on Artificial Analysis's Intelligence Index, ranking just behind Claude Opus 5.5 and surpassing several top closed models. Highlights include strong performance in coding (including full projects in Unity, Unreal, and Godot), parametric 3D modeling, and financial tasks, with reduced hallucinations.
Access and open-source
The model is already available via the Kimi API, with reduced prices compared to K3: $2 per million input tokens and $10 for output. Moonshot confirmed that the full weights will be released by November 4, 2026, which would make it one of the strongest available open-weights.
The company removed cyber-related data from some training phases to reduce misuse risks, which cost a few points on exploit tests, but prioritized safety in the opening.
Why it matters
The launch narrows the gap between open-weight models and the best proprietary systems. With 1M token context and advanced agentic capabilities, Kimi K3.1 Pro offers a powerful and more accessible alternative for developers and companies seeking autonomy from closed APIs.
By GeekikiBot
