Moonshot AI has introduced Kimi K3.1 Pro, a language model based on the K3 architecture with 2.8 trillion parameters. The model features a 1 million token context window and supports both text and image processing. A high-speed version, K3.1 Pro Highspeed, has also been released, offering double to triple the generation speed of the standard model.
The new model demonstrates improved capabilities in programming, financial analysis, and 3D modeling. It can generate 3D models from text or images and write scripts for CAD software. In financial tasks, it processes corporate reports and builds models. The company reports a reduction in hallucination rates from 23.4% to 8.3% in internal benchmarks.
Developers intentionally restricted the model's cyber-vulnerability exploitation capabilities by removing specific security-related data from the training corpus. Despite these limitations, the model scored 81% on the xdev-bench-hard test, compared to 33.3% for its predecessor. Moonshot AI plans to release the full model weights by November 4.
Kimi K3.1 Pro is accessible via the company's website, applications, and API. The API pricing is set at $2 per 1 million input tokens and $10 per 1 million output tokens. While API rates were reduced by one-third, the overall cost per task remains comparable to the previous K3 version.
