DeepSeek: R1 0528

DeepSeek: R1 0528, an open-source 671B parameter model with 37B active in inference, demonstrates strong performance across various benchmarks, often rivaling proprietary models. Its reliability is exceptional, boasting a 99% success rate, indicating consistent and usable responses. However, the model tends to have longer response times, ranking in the 15th percentile for speed, and is positioned at premium pricing levels (18th percentile). In terms of accuracy, DeepSeek: R1 0528 excels in Instruction Following (93rd percentile for one benchmark, though 42nd percentile for another), Coding (80th percentile), General Knowledge (70th percentile), and Reasoning (82nd percentile). Its performance in Ethics (99.0% accuracy) and Mathematics (92.8% accuracy) is also commendable. A notable weakness is its consistently high duration across most benchmarks, with several falling into the lowest percentiles for speed. While its accuracy is high, the associated costs are generally in the lower percentiles, suggesting a premium cost-per-query despite strong performance. The model's fully open reasoning tokens are a significant advantage for transparency and development.