Alibaba has released an updated version of its Qwen 235 billion parameter Mixture-of-Experts (MoE) language model, boasting significant improvements across various domains. This new iteration features 22 billion activated parameters from a total of 128 experts, with eight activated per forward pass. A key innovation is its design for a "non-thinking mode," aiming to overcome reasoning limitations while maintaining core functionalities.
Enhanced Capabilities and User Experience
The model demonstrates substantial advancements in:
- Instruction Following
- Logical Reasoning
- Mathematics
- Coding
- Tool Use
It also offers enhanced long-context understanding up to 256k tokens and a better alignment with user preferences, leading to more helpful and higher-quality text generation. The model exhibits a warmer, friendlier, and more human-like personality in its responses.
Performance Benchmarks and Demonstrations
The video showcases the model's impressive performance through various tests:
- Multilingualism: The model claims to understand and generate text in over 100 languages, with translations appearing accurate (e.g., "I love you" in various languages).
- Literature and Reasoning: It accurately discusses Howard Gardner's "Theory of Multiple Intelligences" and provides structured advice on identifying personal intelligence types.
- Creative Problem Solving: When asked to make durian appealing to everyone, the model generated three innovative solutions: "Scent Stealth Encapsulation Technology" to control aroma, "Personalized Durian Blends via AI Taste Profiling" for customized flavors, and ideas for "Fermented & Gourmet Durian Products." Each solution included detailed appeals and marketing angles.
- Ethical Dilemmas: In response to whether killing a mosquito is morally right, the model provided a balanced, nuanced answer, discussing different ethical perspectives (utilitarian, biocentric) without taking a definitive side.
- Mathematics: It successfully solved a number theory problem (finding the GCD) using two methods and correctly answered another math question. The "thinking mode" can also be enabled to show its reasoning process.
- Code Optimization: The model effectively optimized a complex SQL query, identifying key issues and suggesting improvements like window functions, efficient aggregation, and indexing, presented in a clear tabular format.
- Creative Coding: It generated a self-contained HTML file for a romantic wedding animation with specified elements, animations, and technical specs. While the initial animation had room for improvement, it demonstrated the model's capability to produce complex, standalone code.
- Real-World Social Interaction: When presented with a flirtatious scenario, the model offered respectful and consensual advice, providing polite opening lines and emphasizing the importance of genuine connection and authenticity over manipulative tactics.
Conclusion
This minor update to Qwen 3's MoE model has resulted in significant performance improvements, especially in coding and reasoning, and is seen as a substantial leap forward for open-source language models. The separation of "instruct" and "thinking" modes is highlighted as a strategic move that enhances the model's purpose-built capabilities.
What are your thoughts on Alibaba's latest open-source AI model? Have you tried it out yet?