dots3-note Model Leads Open-Weight Benchmarks with 75.1 Score

Xiaohongshu AI lab releases 280B MoE model dots3-note, scoring 75.1 on Terminal-Bench 2.1. It outperforms peers by 4.9 points with multimodal support and TEMPO RL training.

Woofun AI reports that dots.studio, an AI laboratory, has open-sourced the preview version of its dots3-note model. The model utilizes a 280 billion-parameter Mixture of Experts architecture with 16 billion active parameters and supports a context window of 512,000 tokens. It features multimodal capabilities for text, vision, and audio, incorporating the TEMPO reinforcement learning method for long-term agent training. Model weights are available on Hugging Face, and the API is integrated into OpenRouter.

Data from SemiAnalysis indicates that dots3-note achieved a score of 75.1 on the Terminal-Bench 2.1 benchmark. This result exceeds the previous best-performing open-weight model by 4.9 points. SemiAnalysis noted that their team continues to evaluate the model's practical performance to verify its actual quality and check for potential over-optimization regarding evaluation metrics.