Summary

VibeThinker-1.5B is a 1.5 billion-parameter model demonstrating that small models can achieve robust reasoning capabilities. It targets engineers and researchers seeking highly efficient, cost-effective reasoning solutions. The project delivers state-of-the-art performance in mathematical and coding tasks with a fraction of the parameters and training cost of leading models.

How It Works

The model uses the "Spectrum-to-Signal Principle (SSP)" post-training methodology. This involves "Two-Stage Diversity-Exploring Distillation" in SFT to generate diverse solutions, followed by "MaxEnt-Guided Policy Optimization (MGPO)" in RL to amplify correct reasoning signals. This approach enhances logical deduction abilities in a compact model.

Quick Start & Requirements

Requires transformers>=4.54.0; vLLM==0.10.1 or SGLang>=0.4.9.post6 recommended for inference. Model checkpoints are available via Hugging Face and ModelScope. Evaluation programs and sample responses are prepared. Recommended usage parameters: temperature 0.6/1.0, max token length 40960, top_p 0.95, top_k -1 (for vLLM/SGLang). Direct links to these resources are not provided in the README.

Highlighted Details

• Ultra-Efficient: 1.5B parameters, 100x-600x smaller than models like Kimi K2 (1000B+) and DeepSeek R1 (671B).
• Superior Performance: Outperforms DeepSeek R1 on AIME24, AIME25, HMMT25 benchmarks; rivals MiniMax-M1; surpasses Magistral Medium and Claude Opus 4.
• Cost-Effective Development: Post-training cost of $7,800, a 30x-60x reduction from DeepSeek R1 ($294K) and MiniMax-M1 ($535K).

Maintenance & Community

The README does not specify community channels or provide details on ongoing maintenance or active contributors beyond the paper authors.

Licensing & Compatibility

Licensed under the MIT License, which is generally permissive for commercial use and integration into closed-source projects.

Limitations & Caveats

The model is explicitly recommended for competitive math and coding problems. Its performance or limitations on other task types are not detailed. No information is provided on known bugs or unsupported platforms.