QwQ-Buddy-32B-Alpha

QwQ-Buddy-32B-Alpha

Model Summary

QwQ-Buddy-32B-Alpha is a merged 32B model created by fusing two high-performing models:

• huihui-ai/QwQ-32B-Coder-Fusion-9010 (strong in coding and logical reasoning)
• OpenBuddy/openbuddy-qwq-32b-v24.2-200k (strong in general knowledge and reasoning)

The merge was performed using Spherical Linear Interpolation (SLERP) to ensure a smooth and balanced integration of capabilities from both source models. The result is a powerful and versatile 32B model that excels in both coding and reasoning tasks, making it one of the top candidates for leaderboard evaluations.

Model Details

• Model Type: Merged LLM (Qwen-2.5 32B architecture-based)
• Precision: bfloat16
• Merge Method: SLERP (Spherical Linear Interpolation)
• Weight Type: Original (fully merged model, NOT delta-based)
• Context Length: 200K tokens (inherits capabilities from OpenBuddy-QwQ)
• Training Base Models:
  • huihui-ai/QwQ-32B-Coder-Fusion-9010
  • OpenBuddy/openbuddy-qwq-32b-v24.2-200k
• Merged Layers:
  • 0-32 equally distributed from both models
  • 24-64 optimized for knowledge reasoning and logical computations

Performance Improvements

✅ Stronger coding capabilities (inherits high performance from QwQ-32B-Coder-Fusion-9010) ✅ Enhanced general knowledge & reasoning (boosted by OpenBuddy-QwQ) ✅ Balanced self-attention and MLP layers for smoother response generation ✅ Higher robustness in multilingual support (OpenBuddy-QwQ contributions) ✅ Fine-tuned SLERP weighting for best accuracy in benchmarks

Expected Leaderboard Performance

Based on internal testing and model comparisons, QwQ-Buddy-32B-Alpha is expected to achieve top 20 rankings in:

• HumanEval (coding tasks)
• MMLU (multi-task language understanding)
• HellaSwag (commonsense reasoning)
• BBH (Big Bench Hard) (complex problem-solving)

Limitations & Considerations

• 🚧 Not fine-tuned post-merge (raw merge evaluation may have slight instabilities)
• 🚧 No explicit safety alignment applied (inherits behavior from base models)
• 🚧 Performance on unseen edge cases requires additional evaluation

How to Use

To load the model for inference:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "FINGU-AI/QwQ-Buddy-32B-Alpha"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="bfloat16")

inputs = tokenizer("Write a Python function to compute Fibonacci numbers:", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0]))

Acknowledgments

This model was built using:

• MergeKit for SLERP-based weight interpolation
• Hugging Face Transformers for model loading and testing
• Leaderboard Evaluation Benchmarks for performance comparisons

Contact & Feedback

For any inquiries, issues, or feedback regarding QwQ-Buddy-32B-Alpha, please reach out via GitHub or Hugging Face discussions.