Alibaba Group is poised to release its latest artificial intelligence model, Qwen3, in April 2025. This launch follows the introduction of Qwen2.5 in September 2024 and marks a significant advancement in Alibaba’s AI capabilities.
Qwen3 introduces two primary versions: Qwen3-8B and Qwen3-MoE-15B-A2B. The Qwen3-8B model continues the Qwen series’ strengths in language understanding and generation. The Qwen3-MoE-15B-A2B model employs a Mixture-of-Experts (MoE) architecture, enabling cost-efficient training and improved performance by activating specific subsets of its parameters during inference.
Building upon its predecessors, Qwen3 is equipped with larger datasets, enhanced architectures, and superior fine-tuning, allowing it to handle more complex reasoning, language understanding, and generation tasks.
The Qwen3 family also includes specialized models:
- Qwen3-Coder: An advanced code generation model designed to improve upon Qwen2.5-Coder, offering enhanced reasoning capabilities for developers.
- Qwen3-Audio: A model tailored for audio signal processing, enabling direct textual responses to spoken instructions and supporting voice chat and audio analysis modes.
The integration of Qwen3 into the vLLM inference framework indicates its readiness for deployment in various applications.
As Alibaba continues to enhance its AI offerings, the release of Qwen3 is anticipated to bolster its position in the competitive AI landscape, providing advanced tools for businesses, developers, and researchers.
Last week, Alibaba released Qwen2.5-Omni-7B and Qwen2.5-VL-32B , a smaller model focused on developing cost-effective AI agents
Recently, the company also launched a smaller 32 billion parameter model QwQ-32B (model name :qwq-32b-preview), the previous generation flagship Alibaba’s Qwen 2.5 Max currently ranks 11th.
CometAPI has updated the latest QwQ-32B API(model name :qwq-32b-preview) and Qwen 2.5 Max API (model name: qwen-max) .