On November 13, 2025, Weibo's AI division unveiled its open - source large language model, VibeThinker - 1.5B. This model boasts 1.5 billion parameters and has been fine - tuned based on Alibaba's Qwen2.5 - Math - 1.5B. It is freely accessible on platforms like Hugging Face, GitHub, and ModelScope, allowing researchers and enterprise developers to utilize it, even for commercial use. VibeThinker - 1.5B adopts the 'Spectrum - Signal Principle' (SSP) training framework. This framework divides the processes of supervised fine - tuning and reinforcement learning into two distinct stages. In the first stage, the focus is on diversity. The model is encouraged to explore every possible path for problem - solving. In the second stage, reinforcement learning is employed to optimize the best path, accurately pinpointing the correct reasoning route. When evaluated on math test sets such as AIME24, AIME25, and HMMT25, as well as the LiveCodeBench v6 programming test, this model either outperforms or comes close to models that have tens to hundreds of times more parameters. What's more remarkable is that the entire post - training cost for this model is less than $8,000. This is a stark contrast to the costs associated with similar or larger - scale models, which can run into hundreds of thousands or even millions of dollars.