Dark Side of the Moon releases Kimi K3.1 Pro: 3 trillion parameters, 1 million contexts

Abstract:

On October 24,

Dark Side of the Moon Kimi officially launched the first model of the Kimi K3.1 series - K3.1 Pro early this morning.

According to the official introduction, K3.1 Pro has reached or is close to the strongest closed-source model level in terms of code, agent and knowledge benchmarks. In the Terminal-Bench 4.0 evaluation, K3.1 Pro scored 65.7%, exceeding GPT-6 Astra's 58.2%, and only slightly behind Claude Opus 5.5's 66.4%. In the internal benchmark Kimi Code Bench 3.0 Long-horizon, the K3.1 Pro scored 61.2%, higher than the K3's 55.95%.

In terms of 3D creation and spatial reasoning, K3.1 Pro scored 92.7% with the help of tools in the BenchCAD benchmark test, which is close to Claude Opus 5.5's 96.2% and GPT-6 Astra's 95.9%. The model can infer object shape, proportions and spatial relationships based on text descriptions, sketches or reference drawings, and reconstruct them into usable 3D models. kimi.com and Kimi App have simultaneously launched the 3D Make function, supporting online preview, editing and 3D printing export.

In terms of financial workflow, K3.1 Pro scored 64.4 in the OfficeQA Pro v2 benchmark test, achieving the current best result. On the three internal financial benchmarks, financial modeling, financial research, and business operations have improved by 14%, 20%, and approximately 4.2 times respectively compared with K3. In terms of factual accuracy, in the internal stress test Kimi Hallu Bench Hard, the single fact statement error rate of K3.1 Pro dropped from 23.4% of K3 to 8.3%.

From now on, users can use K3.1 Pro and high-speed versions through kimi.com, Kimi App, Kimi Work, Kimi Code and Kimi API. The API pricing is 13 yuan per million Token input (misses the cache), 1.3 yuan (cache hits), and outputs 68 yuan; the high-speed version is twice as expensive. Full model weights will be released before November 4th.