Model WarsJune 28, 2026via The Decoder
Sina's open model VibeThinker-3B aims to show reasoning compresses well but factual knowledge doesn't
Why it matters
VibeThinker-3B challenges conventional wisdom that model capability requires scale, demonstrating that reasoning can be efficiently compressed through multi-stage post-training while exposing a fundamental tradeoff: logical reasoning vs. factual knowledge retention. This has direct implications for edge deployment and cost optimization strategies.
Key signals
- VibeThinker-3B: 3 billion parameters
- Matches DeepSeek V3.2 and Kimi K2.5 on math and coding benchmarks
- 333x smaller than comparable models
- Multi-stage post-training approach
- Hypothesis: reasoning compresses well, broad world knowledge does not
The hook
3B parameters. 333x smaller than DeepSeek V3.2. Same math and coding scores. Sina's VibeThinker-3B just proved reasoning compresses—but knowledge doesn't.
Sina Weibo's VibeThinker-3B has just three billion parameters but matches models like DeepSeek V3.2 and Kimi K2.5 on math and coding benchmarks. Those models are up to 333 times larger. The secret isn't size but multi-stage post-training. The researchers propose a hypothesis based on their findings: logical reasoning compresses well into small models, but broad world knowledge does not.