Moonshot AI Releases 2.8T Parameter Kimi K3.1 Pro Open-Weight Model
Moonshot AI has released the open-source weights for its 2.8-trillion-parameter Kimi K3.1 Pro model on Hugging Face, claiming the top spot among open-weight models on the Agent Arena leaderboard. The model achieves leading performance across work, coding, and chat categories, backed by evaluations from over 17,500 real-world user agent tasks. The release provides the open-source community with an ultra-large-scale model that rivals leading proprietary systems in agentic workflows and coding benchmarks. This marks a major milestone for open-weight models in handling long-horizon tasks and complex multi-step execution. Kimi K3.1 Pro utilizes a Mixture of Experts architecture with 2.8 trillion total parameters, activating approximately 104 billion per inference, and supports a context window of around 1 million tokens. However, its massive file size requires roughly 1.4TB of storage even with INT4 quantization, making local deployment on standard consumer hardware impractical.
## BACKGROUND
Agent Arena evaluates AI models by extracting signals from real user agent sessions rather than traditional blind tests, estimating net improvement rates using a randomized inference method called causal tracing. Mixture of Experts (MoE) is a neural network architecture that improves performance by routing input data to different sub-models, or experts, instead of activating the entire network.