Moonshot AI Unleashes Kimmy K3.1, Reaching Frontier Performance Amidst Distillation Debate
Moonshot AI has officially launched Kimmy K3.1, a new 3 trillion-parameter open-weight model slated for full download availability on November 4th. The model initially appeared anonymously as ‘Kibitz’ on Chapot Arena, rapidly demonstrating frontier-level performance in coding tasks, a reveal that was quickly attributed to Moonshot due to early naming conventions. Kimmy K3.1’s benchmark results are notably impressive, positioning it within one point of Opus 5.5 on Terminal Bench 4.0 and excelling in S.Marathon, a benchmark for multi-hour software requirement completion. Independent verification by Artificial Analysis further corroborated these high scores, with one test yielding a 67% score, the highest recorded for any model tested by them.
Kimmy K3.1’s breakthrough performance casts a fresh light on the ongoing industry debate surrounding AI distillation. Its advanced capabilities challenge recent assertions by Anthropic and OpenAI that they have largely neutralized large-scale distillation attacks, particularly from Chinese entities. These larger AI companies have long implied that the rapid progression of Chinese models is often tied to unauthorized use of their proprietary outputs and reasoning tokens. Furthermore, Moonshot AI has confirmed a strategic decision to deliberately reduce Kimmy K3.1’s cybersecurity capabilities, lowering its success rate on internal illegal exploit tests from 90.5% to 81%. This move is seen as a pre-emptive measure following the recent US ban of Deepseek models from government use, aiming to mitigate regulatory risks. Despite this self-imposed limitation, Kimmy K3.1 still emerged as the top performer among publicly available models on Artificial Analysis’s new Cyber Index, although competitors like Claude and GPT-6 reportedly declined many prompts on safety grounds. The model also demonstrated robust general coding agent capabilities, including complex 3D rendering tasks, though specific animation quality was noted as an area for further refinement compared to more sophisticated skeletal animation techniques.