VibeThinker: The Big Logic of Small Models—Diversity-Driven Optimization Unleashes Large Model-Level Reasoning Capabilities

Weibo AI's open-source VibeThinker series small models (1.5B/3B) challenge the traditional perception that "only large models have strong reasoning capabilities" with extremely low training costs (1.5B costs only $7,800). Through the innovative Spectrum-to-Signal Principle (SSP) training paradigm, they outperform large models like DeepSeek R1 (with 400x more parameters) on math competitions such as AIME and HMMT, as well as programming tasks, demonstrating the reasoning potential of small models.

VibeThinker was developed by the Weibo AI team. The 1.5B version was first open-sourced in November 2025, and the 3B version was released in June 2026. The core breakthrough of the project lies in achieving reasoning performance that surpasses large models at an extremely low cost (the 1.5B training cost is $7,800, which is 30-60 times lower than DeepSeek R1's $294K), redefining the economics of reasoning models. The base model uses the Qwen2.5-Coder series, leveraging the verifiability of code data to cultivate rigorous reasoning capabilities.

The core innovation of VibeThinker is the SSP training paradigm:

1. Diversity Exploration Distillation: Generate diverse reasoning trajectory "spectra" during the SFT phase to ensure the model covers multiple problem-solving approaches;
2. Signal Amplification: Strengthen correct "signals" from the spectra via MaxEnt-Guided Policy Optimization (MGPO) during the RL phase. The 3B version upgrades the SSP process (enhanced data synthesis, multi-domain RL, long context retention, etc.) and introduces the Claim-Level Reliability Assessment (CLR) strategy, which performs reliability assessment on each claim during reasoning to correct errors, significantly improving accuracy.

Practical performance data verifies the superiority of small models:

• 1.5B version: AIME24 (80.3 vs DeepSeek R1's 79.8), HMMT25 (50.4 vs 41.7), outperforming DeepSeek R1 with 400x more parameters;
• 3B version: AIME26 (94.3→97.1 with CLR), HMMT25 (89.3→95.4 with CLR), LiveCodeBench v6 (80.2 Pass@1), 96.1% acceptance rate on recent LeetCode problems, achieving cutting-edge performance.

Recommended Scenarios: Competition-level math problems (AIME/HMMT), programming competitions (LeetCode/LiveCodeBench), STEM reasoning, instruction-following tasks; Limitations: Not suitable for broad open-domain knowledge tasks; advantages are concentrated on verifiable reasoning tasks. Recommended reasoning configurations: temperature 0.6/1.0, top_p=0.95, top_k=-1, max_tokens=40960.

Open-Source Contributions: Provide a cost-effective reasoning model development path, reference implementation of the SSP paradigm, complete evaluation toolchain, detailed hyperparameter configurations; ranked first on the trending list after release on Hugging Face; Technical Insights: Data quality is better than scale, diversity is key to discovery, focus on optimization of verifiable tasks; Future Directions: Extend SSP to larger models, cross-model application of CLR strategy, multi-domain verifiable task processes, new paradigm of small models + external tools.

VibeThinker, through the SSP paradigm and extremely low cost, proves the great potential of small models in reasoning tasks, challenging the traditional concept that "scale determines capability". It provides a cost-effective path for developing high-performance reasoning models for researchers/developers with limited resources, promoting innovation in the field of small model reasoning.