DeepSeek Open-Sources Prover-V2: Major Boost in Math Reasoning
DeepSeek Prover-V2: Open-source AI model excels in math reasoning, achieving 88.9% on MiniF2F. Supports Lean 4 & 32K context.
"AI Disruption" Publication 6200 Subscriptions 30% Discount Offer Link.
Prelude to DeepSeek R2?
As Labor Day arrives, DeepSeek’s updates keep coming.
Rumors about the imminent release of DeepSeek-R2 have been circulating, but while DeepSeek has indeed made a move, it’s not R2. Instead, we get DeepSeek-Prover-V2, which, of course, is open-source.
Prover-V2 achieves industry-leading performance in theorem proving, with an 88.9% pass rate on the MiniF2F test and strong scores on AIME 24 and 25.
On the evening of April 30, the machine learning platform HuggingFace updated some technical details about DeepSeek-Prover-V2.
The DeepSeek team released two versions of the DeepSeek-Prover-V2 model, with parameter scales of 7B and 671B.
DeepSeek-Prover-V2-671B was trained based on DeepSeek-V3-Base, while DeepSeek-Prover-V2-7B was built on DeepSeek-Prover-V1.5-Base, supporting context length extensions up to 32K tokens.