Paper 2511.06221
Tiny Model, Big Logic: Diversity-Driven Optimization Elicits Large-Model Reasoning Ability in VibeThinker-1.5B
- Published
- Nov 2025
- Research lab
- Independent
- Citations
- 8
- GitHub
- 1.5K stars
01 In brief
Summary
This technical report from Sina Weibo Inc.
introduces VibeThinker-1.5B, a 1.5B-parameter dense model that challenges the assumption that small models lack robust reasoning capabilities.
Developed with a post-training methodology based on the 'Spectrum-to-Signal Principle (SSP)', the approach decouples SFT and RL objectives: the SFT 'Spectrum Phase' uses 'Two-Stage Diversity-Exploring Distillation' to generate a broad spectrum of solutions, while the RL 'Signal Phase' uses 'MaxEnt-Guided Policy Optimization (MGPO)' to amplify correct signals, prioritizing problems where the model's performance is most uncertain.
With a total training cost of $7,800 (3,900 GPU hours on H800), VibeThinker-1.5B surpasses closed-source models like Magistral Medium and Claude Opus 4 on AIME24 and AIME25, and outperforms DeepSeek R1 (671B) on AIME24 (80.3 vs.
79.8), AIME25 (74.4 vs.
70.0), and HMMT25 (50.4 vs.
41.7).
It also achieves 51.1 on LiveCodeBench V6, exceeding Magistral Medium's 50.3.
These results represent substantial improvements over its base model (Qwen2.5-Math-1.5B), which scored 6.7 on AIME24 and 0.0 on LiveCodeBench V6.
The authors argue that small models can achieve reasoning parity with large models, democratizing access to advanced AI research.
The model checkpoint is open-sourced on GitHub and HuggingFace.
02 From the paper
Abstract
Challenging the prevailing consensus that small models inherently lack robust reasoning, this report introduces VibeThinker-1.5B, a 1.5B-parameter dense model developed via our Spectrum-to-Signal Principle (SSP). This challenges the prevailing approach of scaling model parameters to enhance capabilities, as seen in models like DeepSeek R1 (671B) and Kimi k2 (>1T). The SSP framework first employs a Two-Stage Diversity-Exploring Distillation (SFT) to generate a broad spectrum of solutions, followed by MaxEnt-Guided Policy Optimization (RL) to amplify the correct signal. With a total training cost of only $7,800, VibeThinker-1.5B demonstrates superior reasoning capabilities compared to closed-source models like Magistral Medium and Claude Opus 4, and performs on par with open-source models like GPT OSS-20B Medium. Remarkably, it surpasses the 400x larger DeepSeek R1 on three math benchmarks: AIME24 (80.3 vs. 79.8), AIME25 (74.4 vs. 70.0), and HMMT25 (50.4 vs. 41.7). This is a substantial improvement over its base model (6.7, 4.3, and 0.6, respectively). On LiveCodeBench V6, it scores 51.1, outperforming Magistral Medium's 50.3 and its base model's 0.0. These findings demonstrate that small models can achieve reasoning capabilities comparable to large models, drastically reducing training and inference costs and thereby democratizing advanced AI research.