VibeThinker-3B-oQ4

VibeThinker-3B-oQ4

oQ4 quantized version of VibeThinker-3B optimized for Apple Silicon.

This release provides a compact and efficient MLX conversion of VibeThinker-3B for local inference while preserving the strong reasoning, mathematics, and coding capabilities of the original model.

About VibeThinker-3B

VibeThinker-3B is a compact 3B parameter reasoning model developed by WeiboAI to explore the limits of verifiable reasoning in small language models.

Building upon the Spectrum-to-Signal (SSP) training methodology, VibeThinker-3B combines curriculum-based supervised fine-tuning, reinforcement learning, and self-distillation to achieve strong performance across:

• Mathematical reasoning
• Competitive programming
• Code generation
• STEM problem solving
• Instruction following
• Long-form reasoning

Despite its small size, VibeThinker-3B achieves performance competitive with significantly larger reasoning models on several challenging benchmarks. Its development focuses on maximizing reasoning efficiency rather than scaling parameter count alone.

Quantization

This release uses standard oQ quantization.

• Format: MLX
• Quantization: oQ4
• Group Size: 64
• Target Platform: Apple Silicon

This is not an oQ mixed-precision quantization. The model uses standard oQ4 quantization for efficient local deployment.

Recommended Settings

For best reasoning performance:

temp: 0.9
top_p: 0.95
rep_penalty: 1.05
presence_penalty: 1.5
enable_thinking: true

These settings encourage deeper reasoning chains while maintaining output diversity and stable long-form responses.

Example Usage

from mlx_lm import load, generate

model, tokenizer = load("VibeThinker-3B-oQ4")

prompt = """
Solve this problem step by step:

If a train travels 240 km in 3 hours,
what is its average speed?
"""

response = generate(
    model,
    tokenizer,
    prompt=prompt,
    temp=0.9,
    top_p=0.95,
)

print(response)

Intended Use

This model is well suited for:

• Mathematical reasoning
• Programming assistance
• Algorithm design
• Competitive programming practice
• STEM education
• Research experiments
• Local reasoning agents

Hardware Recommendations

Recommended:

• Apple M1 Pro / Max
• Apple M2 Pro / Max / Ultra
• Apple M3 Series
• Apple M4 Series

Compatible with:

• MLX
• oMLX
• Open WebUI
• Custom agent frameworks
• Local Apple Silicon deployments

Credits

Original Model

All credit for the original model, training methodology, datasets, and research belongs to the WeiboAI team.

Original model:

WeiboAI/VibeThinker-3B

Quantized Release

This repository contains only an oQ4 quantized conversion of the original model for efficient Apple Silicon deployment.

Acknowledgements

• WeiboAI
• VibeThinker Research Team
• Apple MLX
• Hugging Face

References

VibeThinker-3B extends the Spectrum-to-Signal training paradigm and demonstrates that frontier-level reasoning capabilities can emerge in compact language models through optimized post-training rather than parameter scaling alone.

Citation

Please cite the original VibeThinker-3B paper when using this model in research.

@article{xu2026vibethinker3b,
  title={VibeThinker-3B: Exploring the Frontier of Verifiable Reasoning in Small Language Models},
  author={Sen Xu and Shixi Liu and Wei Wang and Jixin Min and Yingwei Dai and others},
  year={2026},
  eprint={2606.16140},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}