DeepSeekMath 7B

DeepSeekMath-7B, released February 2024, was the model that announced DeepSeek as a serious lab: a 7-billion-parameter model continued-pretrained on 120B math-heavy tokens that, with chain-of-thought prompting, scored 51.7% on the MATH benchmark — ahead of much larger general models at the time. The instruct and RL variants (DeepSeekMath-Instruct, DeepSeekMath-RL) pushed further with tool-integrated reasoning.

Its real legacy is as a foundation: Project Numina's AIMO-winning NuminaMath models were fine-tuned on DeepSeekMath-7B-Base, and the paper's data recipe (high-quality web math + synthetic reasoning traces) became the template for every open math model after it. The code is MIT-licensed and the model weights carry a DeepSeek license that permits commercial use.

Honest caveats: it is a 2024-era specialist — newer math models (Qwen2.5-Math, NuminaMath, R1 distills) surpass it, and it needs explicit chain-of-thought prompting ('reason step by step, put the final answer within \boxed{}') to perform. Still historically important and a solid lightweight math base.