Chinese artificial intelligence start-up DeepSeek uploaded on Wednesday a new specialized AI model dedicated to solving math problems to an open-source AI community, Hugging Face, without releasing an official statement or posting about it on its social media.

The new model, Prover-V2-671B, does not have an official description on the company's Hugging Face page, but it has 671 billion parameters and is built on top of DeepSeek's V3 model. It represents an update for the company's Prover-V1.5, introduced last year, which was designed for theorem proving in Lean 4.

DeepSeek's new model follows an announcement by Alibaba Group Holding Limited, which launched its latest Qwen 3 series of large language models (LLMs) yesterday, that included the model Qwen3-235B-A22B, dedicated to solving math and coding queries based on two models, Thinking and Non-Thinking.