Moonshot AI Unveils Kimi K3.1 Pro: A 3-Trillion Parameter Powerhouse
China's Moonshot AI unveils Kimi K3.1 Pro, a 3-trillion-parameter model with a 1M token window. Full model weights coming by November 4.
In a bold move to push the boundaries of artificial intelligence, Chinese AI firm Moonshot AI has officially launched Moonshot AI Kimi K3.1 Pro. This new 3-trillion-parameter open-weight model is designed to tackle complex tasks with a massive 1-million-token context window, marking a significant leap in the company's AI capabilities.
Alongside the flagship model, the company has also introduced Kimi K3.1 Pro Highspeed, a variant engineered for those who require rapid-fire responses and more responsive interactions, boasting speeds 2 to 3 times faster than the standard Pro version.
A Massive Leap in Capability
The Kimi K3.1 Pro is not just a minor update; it is an upgraded version of the previous Kimi K3. Moonshot AI revealed that the model was developed using substantially more computing resources during its post-training phase. This investment has resulted in a model that excels in several specialized domains:
- Advanced Coding: From fixing minor bugs in existing codebases to building entire projects from scratch.
- Financial Research: The ability to handle end-to-end delivery of complete financial workflows.
- Creative Work: Enhanced capabilities in 3D creation and other agentic tasks.
Outperforming the Giants?
Moonshot AI claims that K3.1 Pro performs on par with some of the world's leading proprietary AI models. To back this up, the company pointed to specific benchmark results:
On the Terminal-Bench 4.0, which tests coding agents on real-world terminal tasks, Kimi K3.1 Pro scored 65.7 per cent. This outperforms GPT-6 Astra, which scored 58.2 per cent, and puts it in close competition with Claude Opus 5.5, which scored 66.4 per cent.
Additionally, on the company's own Kimi Code Bench 3.0 Long-horizon—a test for multi-step execution and tool use across engineering domains—the K3.1 Pro scored 61.2 per cent, a notable increase over the 55.95 per cent achieved by its predecessor, K3.
Infrastructure and Availability
The performance gains are attributed to an optimized reinforcement learning (RL) framework, improved long-context processing, and enhanced training efficiency. These infrastructure upgrades allow the model to operate effectively across hundreds of application areas, including autonomous AI tasks and knowledge-based operations.
Where to access Kimi K3.1 Pro:
The model and its high-speed variant are currently available through several channels:
- Kimi.ai website
- Kimi mobile applications
- Kimi Work and Kimi Code
- The Kimi API for developers
In a move that will excite the global developer community, Moonshot AI has announced plans to release the full model weights by November 4, further cementing its commitment to the open-weight AI ecosystem.