DeepSeek Unveils Prover V2: A Leap Forward in AI-Powered Mathematical Problem-Solving

DeepSeek Quietly Releases Prover V2

Chinese AI startup DeepSeek has quietly unveiled Prover V2, the latest iteration of its specialized AI system designed for mathematical problem-solving and theorem proving. The company uploaded the new model to the AI development platform Hugging Face on April 30, 2025, without any formal announcement 1

1

5

5

.

Advanced Architecture and Capabilities

Prover V2 is built on DeepSeek's V3 model, boasting an impressive 671 billion parameters and utilizing a mixture-of-experts (MoE) architecture 1

1

3

3

. This architecture allows the model to break down complex tasks into subtasks and delegate them to specialized "expert" components, enhancing its problem-solving capabilities 1

1

.

The new model is designed to translate mathematical problems into formal logic using the Lean 4 programming language, a tool widely used for proving theorems 2

2

. It can solve high-school to college-level mathematical problems, find and fix errors in proofs, generate step-by-step explanations, and assist researchers in exploring new theorems 4

4

.

Open-Source Availability and Versions

DeepSeek has released Prover V2 as an open-source model under the permissive MIT license 2

2

. It is available in two sizes:

1. A 7 billion parameter version built upon DeepSeek-Prover-V1.5-Base
2. A larger 671 billion parameter version trained on top of DeepSeek-V3-Base 4

   4

The model has been quantized to 8-bit floating point precision, effectively halving its size to approximately 650 gigabytes 2

2

.

Training Process and Improvements

The researchers implemented a cold-start training system by prompting the base model to decompose complex problems into subgoals. The proofs of resolved subgoals were then added to the chain-of-thought (CoT) and combined with the reasoning of the base model to create an initial cold start for reinforcement learning 4

4

.

While specific improvements in Prover V2 are yet to be detailed, its predecessor, Prover V1.5, had already shown optimizations in both training and execution, achieving higher accuracy in benchmarks 2

2

.

Implications for AI Development

The release of Prover V2 highlights the ongoing competition in AI development, particularly in specialized domains like mathematics. It follows closely on the heels of Alibaba's launch of its third-generation Qwen family of AI models 5

5

.

The open-sourcing of such advanced models has sparked discussions about democratizing AI access while raising concerns about potential misuse. However, it also demonstrates that open AI is becoming a serious contender to closed AI systems 2

2

.

DeepSeek's Growing Influence

This release comes at a time when DeepSeek is gaining prominence in the AI field. The company recently updated its general-purpose V3 model and is expected to soon release an update to its R1 "reasoning" model 1

1

3

3

. Reports suggest that DeepSeek may be considering raising outside funding for the first time, indicating potential for further growth and development 1

1

.

As AI continues to advance rapidly, specialized models like Prover V2 showcase the potential for AI to make significant contributions in fields requiring complex reasoning and problem-solving skills, potentially revolutionizing mathematical research and education.