Sina open-sources VibeThinker-3B: reasoning can be compressed, factual knowledge cannot
Original title:Sina's open model VibeThinker-3B aims to show reasoning compresses well but factual knowledge doesn't
AI summary
Sina released VibeThinker-3B, a model with only 3B parameters that matches models 200–333 times larger, such as DeepSeek V3.2, on math and coding benchmarks including AIME26, surpasses all models under 20B on LiveCodeBench, and solves 123/128 LeetCode contest problems, beating GPT-5.2 and Kimi K2.5. However, it lags substantially on the knowledge-intensive GPQA-Diamond. The model is based on Alibaba's Qwen2.5-Coder-3B and underwent multi-stage post-training including SFT, reinforcement learning and self-distillation. The research proposes a "parameter compression-coverage hypothesis": logical reasoning relies on a small number of compressible patterns, while broad world knowledge still requires large parameters. The model has been open-sourced.
Key Takeaways
- Sina released VibeThinker-3B,a model with only 3B parameters that matches models 200–333 times…
- 2,on math and coding benchmarks including AIME26,surpasses all models under 20B on…
- 2 and Kimi K2.
- However,it lags substantially on the knowledge-intensive GPQA-Diamond.
- The model is based on Alibaba's Qwen2.
Quick Facts
- Published
- Updated
- Source
- The Decoder:AI News(RSS)
- Category
- Model Releases
- AI score
- 70
VibeThinker-3B 用 3B 参数在数学编程上匹敌百倍大模型,推理可压缩而知识不能的假设值得深思。对做推理应用的人来说是个信号。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: Sina open-sources VibeThinker-3B: reasoning can be compressed, factual knowledge cannot. June 28, 2026. https://new.nat6.net/en/news/cmqxiel9h004ysl50umg2c011/