DeepSeek’s newly upgraded maths-focused artificial intelligence (AI) model has fuelled public anticipation for the Chinese start-up’s upcoming R2 reasoning model, which remains shrouded in secrecy even as major Chinese tech firms raise the stakes with recent releases of their own large language models.
On Wednesday, DeepSeek quietly open-sourced a new 671-billion-parameter AI model called Prover-V2, an upgrade to its specialised model designed to handle maths proofs. The incremental update of the product, based on its V3 foundational model, highlights DeepSeek’s ability to develop powerful yet relatively small models while dealing with limited access to the world’s most advanced AI chips from Nvidia.
The release captured attention online, as people have been closely watching for updates from the hotshot start-up, but it was not the next-generation leap that many have been waiting for.
The surprise release came just a day after Alibaba Group Holding, owner of the Post, unveiled its Qwen3 family of AI models, which the tech giant said surpassed DeepSeek’s R1 reasoning model in multiple areas. DeepSeek-R1, launched in January, stunned the global AI community with its efficient use of resources, matching OpenAI’s o1 performance at a fraction of the training cost.
OpenAI launched o3 and o4-mini two weeks ago, describing them as its “smartest and most capable models to date”.
DeepSeek has not disclosed a release schedule or details about R2. However, the Hangzhou-based firm’s new Prover model has drawn plaudits from the AI and mathematics communities. One user on X described themselves as “impressed as a math Olympiad student”.
