Kimi K3.1 Pro finally released: the strongest open-source cybersecurity model

Beating AI News Flash: Moonshot AI releases Kimi K3.1 Pro, the first Kimi model with "Pro" in its name. It retains the 2.8 trillion parameter foundation of K3 and primarily enhances Agent capabilities through large-scale reinforcement learning.

On the third-party Artificial Analysis cyber defense index, K3.1 Pro scored 64%, 23 percentage points higher than K3, surpassing DeepSeek V4.1 Flash's 57% to rank first among open-source models. Only OpenAI's non-public GPT-6 Sol (Daybreak Blue) scored higher on the entire leaderboard, at 69%.

Overall capability also improved significantly. Artificial Analysis gave it 55 points, an 11-point increase over K3, ranking fourth globally and just 3 points behind leader Claude Opus 5.5. On Terminal-Bench 4.0, the terminal coding task pass rate jumped from K3's 13% to 67%, taking the top score. 3D capability showed some divergence: it scored 92.7% on the official BenchCAD, close to Opus 5.5's 96.2%; but on the independent interactive 3D evaluation AA-3D, early performance was not as outstanding as coding.

K3.1 Pro's API token unit price is one-third lower than K3, but Artificial Analysis estimates that its average per-question cost on the cyber defense evaluation rose from $2.45 to $2.81, an increase of about 15%. The main reason is that input token consumption roughly doubled, offsetting the price cut.

The model is currently live on Kimi, Work, Code, and API, with full weights planned to be released no later than November 4.

Moonshot AI also proactively removed some cyber attack and defense training data to reduce the risk of misuse after open-weight release. In the official internal vulnerability exploitation test, the control version without risk mitigation measures had a success rate of 90.5%, while the final version was 81%.