Qwen3 30B-A3B MoE — In-Depth LOCAL Testing! (Think & No-Think)



AI Summary

Summary of Quen 330B A3B Model Testing

  1. Introduction
    • Discusses the use of $25 seed funding for a project.
    • Focus on Quen 330B A3B model from Quen 3 family.
  2. Model Comparison
    • Quen 330B A3B with 3 billion parameters outperforms QWQ32B with 32 billion parameters.
    • Currently downloading a larger model (235 billion parameters).
  3. Sampling Parameters
    • Adjustments needed for correct sampling parameters in LM Studio, consulted from Hugging Face page.
    • Parameters:
      • Temperature: 6
      • Top P: 0.95
      • Top K: 20
      • Min P: 0
  4. Model Testing
    • Synthwave Python Game
      • Game created without external dependencies, showing creativity under constraints.
      • Uses movement keys (A, D) for gameplay.
    • Pygame Version
      • Enhanced game with smoother controls and scoring.
    • Website Mockup
      • Generated HTML for fictional PC repair business with a contact form and visual elements, including mobile responsiveness.
    • Testing output differences between thinking and non-thinking modes.
      • Non-thinking mode outputs quicker and emphasizes simplicity.
  5. Conclusion
    • Impressive performance of Quen 330B A3B model.
    • Reminder to check suggested sampling parameters when using the model.
    • Upcoming video expected on the 235 billion parameter model.