README.md

metadata

title: VibeThinker-1.5B Competitive Coding Assistant
emoji: 🧠
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
pinned: false
license: mit

🧠 VibeThinker-1.5B Competitive Coding Assistant

An interactive demo of VibeThinker-1.5B optimized for competitive programming challenges.

⚡ Performance Highlights

• AIME24: 80.3 (surpasses DeepSeek R1's 79.8)
• AIME25: 74.4 (vs DeepSeek R1's 70.0)
• LiveCodeBench V6: 51.1 (competitive coding)
• Training Cost: Only $7,800 USD
• Parameters: 1.5B (400× smaller than DeepSeek R1)

🎯 What It's Best At

✅ Competitive Programming: LeetCode, Codeforces, AtCoder-style algorithm problems
✅ Python Coding Challenges: Problems with clear input/output specifications
✅ Mathematical Reasoning: Complex proofs and formal reasoning tasks
✅ Algorithm Design: Dynamic programming, graph algorithms, optimization problems

⚠️ Important Limitations

This model is specialized for competitive programming, not general software development:

❌ Not suitable for: Building applications, debugging real codebases, using specific libraries
❌ Limited knowledge: Low encyclopedic knowledge, Python-focused training
❌ Overthinking tendency: May generate verbose reasoning for simple tasks
❌ Narrow scope: Optimized for benchmark-style problems, not production code

See community feedback analysis for detailed real-world testing insights

🚀 Features

• 🧠 Intelligent Parsing: Automatic separation of reasoning and solution
• 📊 Token Tracking: Real-time stats on generation time and token usage
• 💻 Clean Code Display: Syntax-highlighted, copyable/downloadable code blocks
• 📱 Responsive Design: Modern UI with collapsible reasoning sections
• 🎨 High Contrast: Readable output with dark code blocks on white background
• 🔄 Loop Detection: Automatically detects and truncates repetitive output

🛠️ Technical Details

Model Information

• Base Model: Qwen2.5-Math-1.5B
• Training Method: Spectrum-to-Signal Principle (SSP)
  • Supervised Fine-Tuning (SFT) for solution diversity
  • Reinforcement Learning (RL) for correct reasoning paths
• Inference Engine: Standard transformers library (PyTorch)
• Token Efficiency: Configurable thinking depth via prompt hints

Hardware Requirements

• Recommended: Nvidia T4 - small (16 GB VRAM)
• Memory Usage: ~3-4 GB VRAM (1.5B params in float16)
• Cost: $0.40/hour on HuggingFace Spaces

Implementation

# Clean, simple transformers implementation
- torch.float16 for efficiency
- device_map="auto" for automatic GPU placement
- Repetition penalty (1.1) to reduce loops
- Automatic loop detection and truncation

📖 Usage Tips

For Best Results:

1. Frame problems competitively: Clear input/output, edge cases, constraints
2. Adjust thinking tokens:
   • 1024-2048 for quick, simple problems
   • 3072-4096 for standard algorithm challenges
   • 6144-8192 for complex multi-step reasoning
3. Use Python: Model trained primarily on Python code
4. Specify format: Request specific output format (function, class, test cases)

Example Prompts:

✅ Good: "Write a function to find the longest increasing subsequence. 
         Include time/space complexity analysis and test with [10,9,2,5,3,7,101,18]"

✅ Good: "Implement Dijkstra's algorithm with a min-heap. Handle disconnected graphs."

❌ Poor: "Debug my React app" (not its purpose)
❌ Poor: "How do I use pandas?" (limited library knowledge)

🔗 Resources

• Model: WeiboAI/VibeThinker-1.5B
• Paper: arXiv:2511.06221
• GitHub: WeiboAI/VibeThinker
• License: MIT

🙏 Credits

Developed by WeiboAI. This Space demonstrates the model with a clean interface and enhanced user experience.

📝 Citation

@article{vibethinker2025,
  title={Tiny Model, Big Logic: Diversity-Driven Optimization Elicits Large-Model Reasoning Ability in VibeThinker-1.5B},
  author={WeiboAI Team},
  journal={arXiv preprint arXiv:2511.06221},
  year={2025}
}