- Chinese tech giant Alibaba has unveiled its new Qwen 3 family of open-weight AI models, introducing eight new variants.
- The flagship Qwen3-235B-A22B MoE model boasts 235 billion total parameters with 22 billion activated.
- Meanwhile, the more compact Qwen3-30B-A3B MoE model – operating with just 3 billion activated parameters – significantly outperforms GPT-4o.
Chinese tech giant Alibaba has launched eight new open-weight AI models in its Qwen 3 series, featuring two innovative Mixture-of-Experts (MoE) architectures. The lineup includes:
- Flagship Model:
- Qwen3-235B-A22B (235 billion total parameters, 22 billion activated)
- Compact MoE Model:
- Qwen3-30B-A3B (30 billion total parameters, 3 billion activated)
The series also introduces six dense models:
- Qwen3-32B
- Qwen3-14B
- Qwen3-8B
- Qwen3-4B
- Qwen3-1.7B
- Qwen3-0.6B
This comprehensive release demonstrates Alibaba’s commitment to providing scalable AI solutions across different computational requirements.
All Qwen 3 models support Hybrid Thinking Modes, enabling them to function as both reasoning-based AI models and traditional LLMs. In Thinking Mode, the model processes information step-by-step for complex reasoning, while Non-Thinking Mode delivers fast, direct responses.
Additionally, Qwen 3 supports 119+ languages and dialects, making it one of the most multilingual AI models available. Alibaba has also enhanced MCP (Multi-Channel Processing) support, further improving the models’ agentic capabilities.
In terms of performance:
- The flagship Qwen3-235B-A22B competes with top-tier models like DeepSeek R1, Grok-3 Beta, Gemini 2.5 Pro, and OpenAI o1.
- Surprisingly, the smaller Qwen3-30B-A3B (with just 3B activated parameters) outperforms DeepSeek V3 and OpenAI’s GPT-4o in benchmarks.
Alibaba highlights Qwen 3’s strong performance in coding, math, science, and general tasks, positioning it as a frontier AI model family from China. With DeepSeek R2 on the horizon, China is emerging as a formidable competitor to Western AI labs.