On October 24, Farsight reported that Moonshot AI has launched the Kimi K3.1 Pro, the inaugural model within the K3.1 lineup. According to official announcements, this model is equipped with an impressive 3 trillion-level parameters, accommodates a 1 million-Token context window, and offers native visual comprehension capabilities. It is specifically designed to bolster functionalities in coding, 3D creation, financial research, and modeling. Additionally, the simultaneously introduced K3.1 Pro Highspeed is a streamlined variant that retains identical model capabilities while delivering an output speed approximately 2 to 3 times faster than its standard counterpart.

The K3.1 Pro model has undergone post-training based on the Kimi K3 framework, with an augmented computational investment aimed at refining multi-step execution and tool-calling proficiencies for complex tasks.

Official tests reveal that the K3.1 Pro achieved a 65.7% score on Terminal-Bench 4.0, a benchmark for evaluating terminal programming tasks. In the company's internal long-form programming assessments, its performance surged from 55.95% with the K3 model to 61.2%.

Another focal point of this update is 3D creation. As per official descriptions, the model is capable of generating 3D models from textual descriptions, sketches, or reference images. It can also modify assets by writing and executing scripts to invoke Blender and CAD tools. The accompanying 3D Make feature is now accessible on the Kimi web platform and App, enabling users to preview, edit, and download model files for 3D printing online.

Developers have the option to access both models via the Kimi API. For the kimi-k3.1-pro model, the pricing is set at 1.3 RMB (with cache hit) and 13 RMB (without cache hit) per million Token input, with an output fee of 68 RMB. The high-speed variant, kimi-k3.1-pro-highspeed, which boasts an output speed approximately 2 to 3 times faster, is priced at double the standard rate. The Kimi Code documentation also specifies that the K3.1 Pro series exclusively supports thinking mode, and any requests with thinking disabled will be declined.

Presently, both models are accessible through the Kimi web platform, App, Kimi Work, Kimi Code, and API. Moonshot AI has announced that the complete model weights will be made available by November 4.