Skip to content
Sen Xu
All research

VibeThinker-1.5B: Tiny model, big logic

Can a compact language model develop reasoning ability usually associated with much larger systems? VibeThinker-1.5B explores that question through diversity-driven post-training.

Read the paper View the code
Benchmark comparison between VibeThinker-1.5B and larger reasoning models
VibeThinker-1.5B reasoning benchmark comparison.

The question behind the project

Reasoning performance is often associated with model scale. Our work asks whether targeted post-training and greater diversity in optimization data can help a much smaller model acquire stronger reasoning behavior.

Why this is worth sharing

Beyond the benchmark results, the project reflects a broader research direction: capability gains do not have to come only from increasing parameter count. Data design and post-training strategy can be equally important parts of the system.

Paper and code

The paper and open-source implementation contain the complete method, experiments, and evaluation.

Read the paper View the code