Qwen 3 NEW Powerful Opensource Hybrid LLM! Beats Deepseek R1 (Fully Tested)
AI Summary
Quen 3 Model Release Overview
- New Models Released:
- Quen 3 series includes:
- 235 billion parameter model (22 billion active parameters)
- 30 billion parameter lightweight model (3 billion active parameters)
- Additional six dense models (0.6 to 32 billion parameters), released under Apache 2.0 license.
- Performance:
- The flagship 235 billion model competes with Deepseek R1, Gro 3, Gemini 2.5 Pro, and OpenAI models, excelling in coding, math, and reasoning benchmarks.
- The 30 billion model shows strong performance for local use against models like GP4 Omni.
- Model Architecture:
- Utilizes mixture of experts architecture with 10% active parameters, reducing training and inference costs.
- Supports 119 languages, pre-trained on 36 trillion tokens with improved reinforcement learning and coding capabilities.
- Efficiency Gains:
- Positioned for scalable AI deployment.
- Getting Started:
- Models available on Model Scope and via Quen’s chatbot.
- Local installation is possible with the released dense models.
- Benchmark Testing:
- Demonstrated capabilities through various tasks, including creating a front-end application and implementing Conway’s Game of Life.
- Overall successful performance in coding tasks, reading comprehension, and logical reasoning.
- Conclusion:
- Quen 3 models provide competitive alternatives to established models like O3 and Deepseek R1 while being open source and efficient.
Key Takeaway
- Recommendation: Users are encouraged to try the Quen 3 models locally as a robust, open-source alternative.