A new development in the AI field has gained attention after researchers introduced a compact system called VibeThinker-3B. This new AI Model was developed by a team of nine researchers at Sina Weibo. It is designed to perform advanced reasoning tasks despite its small size.
The AI Model reportedly matches or even surpasses much larger systems in several benchmarks. These include models from Google DeepMind, OpenAI, Anthropic, and DeepSeek. The results have sparked interest across the global AI community.
VibeThinker-3B has 3 billion parameters. It achieved a score of 94.3 on AIME 2026. This performance is close to much larger models like DeepSeek V3.2, which has 671 billion parameters. It also outperformed Gemini 3 Pro in the same test.
When enhanced with a reliability method, the AI Model reached an even higher score of 97.1. This improvement highlights its strong reasoning ability in mathematical and structured problem-solving tasks.
In other benchmarks, the model performed strongly across multiple categories. It scored over 90 in several math and coding tests. It also showed high accuracy in instruction-following evaluations.
The AI Model achieved an 80.2 score in LiveCodeBench coding tests. It also performed well in LeetCode challenges, solving 123 out of 128 problems on the first attempt. This result placed it ahead of several leading AI systems in controlled tests.
Despite its performance, VibeThinker-3B is much smaller than competing models. It has around 224 times fewer parameters than some large systems. Researchers noted that it can even run on a consumer laptop.
The team explained that this AI Model focuses on verifiable reasoning tasks. These include mathematics and coding problems. They argue that such tasks can be compressed into smaller models more effectively than general knowledge tasks.
However, the model is not perfect. It performed lower on broad knowledge benchmarks compared to large systems. This shows that the AI Model still has limitations in general understanding.
The training process involved multiple stages. It included supervised learning, reinforcement learning, and final fine-tuning. The model was also trained using a large context window to improve reasoning depth.
Researchers released the model under an open-source MIT License. It is now available on platforms like Hugging Face and ModelScope. Developers have already started creating modified versions.
Experts say this AI Model could support a hybrid AI future. In such systems, small models handle reasoning tasks while large models provide general knowledge.
While the results are impressive, some concerns remain. Critics argue that benchmark performance may not fully reflect real-world performance. More testing is needed in practical software environments.
In other news read more about Apple Intelligence to Support Third-Party AI Models in Major Update
Still, VibeThinker-3B shows how smaller AI Model systems can deliver strong performance. It may influence future research focused on efficient and low-cost AI development.
