China's DeepSeek company recently revealed the launch of an updated version of its model specialized in solving complex mathematical problems, under the name Prover-V2.
This update was released without significant media fanfare, yet it quickly garnered attention across technical platforms. This was particularly notable as it represents a tangible advancement in the capabilities of relatively smaller models, especially given the restrictions imposed on the company regarding access to the latest advanced Nvidia chips.
The new version is based on the foundational V3 model and contains 671 billion parameters. Furthermore, it adopts an architecture following the principle of multiple experts, which allows complex tasks to be divided into smaller parts that are distributed among specialized units.
This approach significantly enhances the model's ability to process mathematical proofs efficiently. Consequently, it has garnered praise from the scientific community; one mathematics olympiad student, for instance, described it as "impressive" on the X platform.
The release of Prover-V2 occurred amid a rising tide of competition among China's tech giants. Notably, it followed just one day after Alibaba Group introduced its Qwen3 series of models, asserting that these models surpassed DeepSeek's earlier R1 model performance in several areas.
The R1 model had achieved significant recognition since its January launch, demonstrating strong capabilities in efficient reasoning while incurring low training costs compared to models such as OpenAI's o1.
On the other hand, OpenAI recently introduced its o3 and o4-mini models, touting them as their smartest and most advanced to date. This move puts additional pressure on DeepSeek to reveal details about its forthcoming R2 model, the specifics of which are currently still under wraps.
Despite the lack of any official release schedule for R2, some analysts believe the introduction of Prover-V2 could serve as an early signal of an impending transition towards the company's next generation of reasoning models.
The updated version of Prover is now available to developers on the Hugging Face platform.
This move underscores DeepSeek's ongoing commitment to providing open models suitable for research, education, and development purposes, while firmly maintaining its focus on AI tailored for mathematical theorem proving.
