DeepSeek Prover V2, an Open-Source Mathematics-Focused AI Model, Released

DeepSeek, the Hangzhou, China-based synthetic intelligence (AI) agency, launched an up to date model of its Prover mannequin on Wednesday. Dubbed DeepSeek-Prover-V2, it’s a extremely specialised mannequin that focuses on proving formal mathematical theorems. The massive language mannequin (LLM) makes use of the Lean 4 programming language to verify if the mathematical proofs are logically constant by analysing every step independently. Similar to the Chinese agency’s earlier releases, the DeepSeek-Prover-V2 is an open-source mannequin and will be downloaded from well-liked repositories akin to GitHub and Hugging Face.

DeepSeek’s New Mathematics-Focused AI Model Is Here

The AI agency detailed the brand new mannequin on its GitHub itemizing web page. It is actually a reasoning-focused mannequin with a visual chain-of-thought (CoT), which features within the area of arithmetic. It is constructed on and distilled from the DeepSeek-V3 AI mannequin, which was launched in December 2024.

DeepSeek-Prover-V2 can be utilized in quite a lot of methods. It can remedy high-school to college-level mathematical issues and discover and repair errors in mathematical theorem proofs. It will also be used as a instructing support and generate step-by-step explanations for proofs, and it could possibly help mathematicians and researchers in exploring new theorems and proving their validity.

It is on the market in two mannequin sizes — a seven billion parameter dimension and a bigger 671 billion parameter dimension. While the latter is educated on prime of DeepSeek-V3-Base, the previous is constructed upon DeepSeek-Prover-V1.5-Base and comes with a context size of as much as 32,000 tokens.

Coming to the pre-training processes, the researchers carried out a cold-start coaching system by prompting the bottom mannequin to decompose complicated issues. These issues served as a sequence of subgoals. Then, the proofs of resolved subgoals had been added to the CoT and mixed with the reasoning of the bottom mannequin to create an preliminary chilly begin for reinforcement studying.

Notably, other than GitHub, the AI mannequin will also be downloaded from DeepSeek’s Hugging Face itemizing. The Prover-V2 mannequin highlights how iterative modifications to the coaching strategy of AI fashions may end up in considerably bettering their specialised functionality. Similar to different open-source mannequin releases, the small print concerning the core structure or the bigger dataset aren’t identified.