Sina Weibo released the VibeThinker-3B, boasting 3 billion parameters that rival flagship models.

According to CryptoBriefing, Sina Weibo's research team has released the VibeThinker-3B language model. This model, with only 3 billion parameters, achieved a score of 94.3 in the AIME 2026 math competition benchmark, comparable to DeepSeek V3.2 with 671 billion parameters. It also achieved a Pass@1 score of 80.2 in the LiveCodeBench v6 coding test. Built on the Qwen2.5-Coder-3B architecture, the model's performance was improved through supervised fine-tuning, multi-domain reinforcement learning, and offline self-distillation techniques. The model weights and code are now fully open-sourced on Hugging Face and GitHub under the MIT license. Analysis indicates that such efficient, small models are easier to run on distributed networks and have constructive implications for decentralized AI projects in the cryptocurrency field.