Chinese artificial intelligence company Moonshot AI has announced the launch of the newest member of its model family, the Kimi K3.1 Pro model, alongside a faster version known as Highspeed. Access to the new model and its fast version has been made available through Kimi's products and the company's Application Programming Interface (API), with plans to officially release the model weights by November 4th.
What does the Kimi K3.1 Pro model practically offer?
The new model was built as an evolution of the Kimi K3 model, with its additional training phase consuming greater computational power. The release comes equipped with a context window of up to one million tokens, allowing it to process extremely long text contents all at once. According to company data, the Highspeed version operates at a speed ranging from two to three times faster compared to the standard model.
The model's new capabilities focus primarily on the fields of software development and 3D design. The company explains that K3.1 Pro is not limited to debugging existing code, but extends to supporting game development projects using popular engines such as Unity, Unreal Engine, and Godot. It can also generate 3D models based on texts, sketches, or reference images, alongside writing code for Blender software and Computer-Aided Design (CAD) to execute required modifications.
Performance and Financial Research
In performance benchmarks, K3.1 Pro achieved a success rate of 65.7% in the Terminal-Bench 4.0 evaluation, which measures command-line interface programming tasks, surpassing some competing models in the results published by the company.
In the field of financial research, the company worked on improving the model's capabilities in reviewing financial documents, comparing data from multiple sources, and performing multi-step calculations. The company's internal validation tests indicated a drop in the rate of hallucination or misinformation based on claims from 23.4% in the previous version to 8.3%, while emphasizing that this rate does not necessarily reflect the general daily error rate.
What does this mean for developers?
The company intends to release the model weights to help developers run and customize it on their own infrastructure. Prior to the release step, the company removed cybersecurity-related content from the training data and reinforcement learning environments in order to restrict capabilities for exploiting vulnerabilities or cyberattacks.