Moonshot AI launches Kimi K3.1 Pro, post-entry version of K3 with a context of a million tokens, prices down by a third and weights promised by November 4.
Moonshot AI has presented Kimi K3.1 Pro, the first model in its K3.1 series. It keeps the K3 architecture (2,800 billion total parameters, 104 billion active, text and images, context of a million tokens) but has received much more post-workout computation, including reinforcement learning conducted in parallel on hundreds of domains. A K3.1 Pro Highspeed version, two to three times faster, accompanies it. Both are available now on Kimi.ai, Kimi Work, Kimi Code and the API; the full weights must be published by November 4.
Just behind the best closed models
On Artificial Analysis' Intelligence Index, K3.1 Pro scores 55 points, eleven more than K3, and ranks fourth behind Claude Opus 5.5 (58), Claude Sonnet 5.5, and Muse Wonder Full (56 each). If the weights come out as expected, it would become the highest-rated open model, and the gap between open and proprietary would drop from twelve to three points. Moonshot claims 65.7 percent on Terminal-Bench 4.0, compared to 58.2 percent for GPT-6 Astra and 66.4 percent for Opus 5.5, and 92.7 percent with tools on BenchCAD, a 3D reconstruction test by CAD code generation. On its own hallucination testbed, the rate of erroneous claims drops from 23.4 percent to 8.3 percent, still above the 5.5 percent of GPT-6 Astra.
The price drops by a third: $2 per million tokens in entry and $10 in output. Moonshot also devotes part of its post to cybersecurity: noting that vulnerability scanning capabilities increased with post-training, the company claims to have removed cyber-related data from the two training phases of the open-weight version, to limit malicious use.
Read the full article on the original source
Go to the source →