DeepSeek: DeepSeek V3.1 Base

DeepSeek-V3.1 Base, a 671B parameter open MoE model with 37B active parameters, demonstrates exceptional speed and cost efficiency. It consistently ranks among the fastest models and offers among the most competitive pricing across all evaluated benchmarks. As a base model, it is designed for raw next-token prediction rather than instruction following. This is evident in its performance, particularly its 0.0% accuracy in Instruction Following. The model shows notable strengths in Coding, achieving 58.3% accuracy, and exhibits strong cost efficiency across most benchmarks, particularly in General Knowledge, Ethics, and Coding. However, its accuracy is generally low across knowledge-based and reasoning tasks. It scored 36.4% in General Knowledge, 50.0% in Ethics, and a mere 2.2% in Reasoning. Mathematics and Email Classification also presented significant challenges, with accuracies of 11.7% and 33.3% respectively. Its performance profile suggests it is best suited for tasks where raw text generation and pattern recognition are paramount, rather than complex instruction adherence or deep reasoning.