DeepSeek: DeepSeek V3 Base

DeepSeek V3 Base, a 671B parameter open Mixture-of-Experts (MoE) model, demonstrates exceptional speed and competitive pricing. It consistently ranks among the fastest models and offers highly competitive pricing across all evaluated benchmarks. Created on March 29, 2025, with a substantial context length of 163840, this base model is designed for testing and requires detailed prompts for optimal performance. In terms of benchmark performance, DeepSeek V3 Base exhibits varying capabilities. Its highest accuracy was observed in Coding (24.1%) and General Knowledge (17.6%), placing it in the 18th and 13th percentile respectively. However, performance in Ethics (13.0% accuracy) and Email Classification (10.0% accuracy) was notably lower, ranking in the 11th and 4th percentile. A significant weakness is evident in Instruction Following, where it achieved 0.0% accuracy. While its cost efficiency is generally good, particularly in General Knowledge and Ethics, its duration for tasks like Ethics and Email Classification is quite high, indicating slower processing times despite its overall speed ranking. This suggests that while the model is fast, its accuracy on certain tasks needs improvement.