The AI Chinese laboratory Deepseek has silently updated Prover, its AI model which is designed to resolve the evidence and theorems related to mathematics.
According to South China Morning Post, Deepseek downloaded the latest version of Prover, V2, on the AI Dev platform, Hugging Face, late on Wednesday. It seems to be built above the V3 model of the startup, which has 671 billion parameters and adopts an architecture for mixing experts (MOE).
The parameters correspond roughly to the problem solving skills of a model, while the MOE breaks down the tasks into subtiles and the delegates to smaller and specialized “experts” components.
Deepseek for the latest prover updated in August, describing it at the time as a personalized model for proving formal theorem and mathematical reasoning.
In February, Reuters reported that Deepseek planned to raise external funds for the first time. The company recently published an improved version of V3, a model for general use, and should soon update its model of “reasoning” R1.
