Kimi K3.1 Pro Finally Released: The Strongest Open Source Cybersecurity Model Dynamic Beating AI News: The Dark Side of the Moon has released the Kimi K3.1 Pro, the first Kimi model with a Pro name. It uses K3's 2.8 trillion parameter base and mainly enhances Agent capabilities through large-scale reinforcement learning. In the third-party Artificial Analysis's network defense index, K3.1 Pro scored 64%, 23 percentage points higher than K3 and surpassing DeepSeek V4.1 Flash's 57%, ranking first among open source models. Only GPT-6 Sol (Daybreak Blue), which OpenAI is not open to the public, has a higher score of 69% on the entire list. The overall ability has also made significant progress. Artificial Analysis gives a score of 55, an increase of 11 points from K3, ranking fourth globally, only 3 points behind the top spot Claude Opus 5.5. In Terminal Punch 4.0, the pass rate of terminal programming tasks jumped from 13% in K3 to 67%, achieving the highest score. 3D capabilities are somewhat differentiated: the official BenchCAD score is 92.7%, close to 96.2% of Opus 5.5; But the independent interactive 3D evaluation of AA-3D's early performance was not as outstanding as programming. The API Token unit price of K3.1 Pro is one-third lower than K3, but Artificial Analysis estimates that the average cost per question for its network defense evaluation has increased from $2.45 to $2.81, an increase of about 15%. The main reason is that the consumption of input tokens has doubled, offsetting the price reduction. At present, the model has been launched on Kimi, Work, Code, and API, and the complete weight plan will be available no later than November 4th. The dark side of the moon also proactively removed some network attack and defense training data to reduce the risk of abuse after opening up weights. In the official internal vulnerability exploitation test, the success rate of the control version without risk reduction measures was 90.5%, and the final version was 81%. [Original link]