Abstract:
Moonshot AI announced that it has officially opened the Kimi K3.1 Pro model weights on the Hugging Face platform. According to the official disclosure of AI model evaluation platform Arena, Kimi K3.1 Pro (Max) ranks first in the open weight model in the Agent Arena agent rankings, 13th among all participating models in the world, and ranks first in the open model in the three major categories of work, programming and chat.
Arena stated that this ranking is based on more than 17,500 real, long-term agent task sessions by global users. The net improvement rate of Kimi K3.1 Pro (Max) reached +5.1%, an increase of 4.2 percentage points from +0.9% of the previous generation Kimi K3 (Max), showing that it has made significant progress in complex task execution and user interaction.
In terms of specific evaluation indicators, many performances of Kimi K3.1 Pro have been improved. Among them, the Bash Reliability indicator, which measures the reliability of Shell command execution, reached +7.9%, ranking first in the open model, while the previous generation K3 was -4.5%, ranking 13th.
In terms of the Praise vs Complaint indicator that reflects user satisfaction, the new model achieved +7.7%, significantly improved from the +1.3% of the previous generation. In terms of the Steerability indicator that measures the model's ability to adjust execution direction based on user feedback, K3.1 Pro achieved +6.5%, also ranking first among open models.
In addition, in the Confirmed Success indicator, which measures the user's clear confirmation of task completion, K3.1 Pro scored +3.9%, ranking fourth among the open models, slightly higher than the previous generation's +3.8%.
According to the model information released by Dark Side of the Moon, Kimi K3.1 Pro adopts a hybrid expert (MoE) architecture with a total parameter volume of 2.8 trillion, a single inference activation parameter of approximately 104 billion, and supports approximately 1 million Token context windows. The model has native multi-modal capabilities, focusing on improving programming, 3D creation, financial research, factual accuracy, and complex agent task execution capabilities.
Dark Side of the Moon stated that the model weights of Kimi K3.1 Pro are now available for download on Hugging Face and will continue to promote the development of an open smart ecosystem in the future.
Comments