Open-Source Multimodal Agent Qwen3.5 Released
A new large-scale, open-source AI agent named Qwen3.5 has been released, featuring 397 billion parameters and a Mixture-of-Experts architecture for efficient inference. The model has native multimodal capabilities, allowing it to reason across text and images, and supports over 200 languages. Its open weights make it available for developers to build upon for consumer and SaaS applications requiring global reach or cross-media automation.
- The Qwen series of models is developed by Alibaba Cloud, with "Qwen" translating to "a thousand questions". The latest iteration, Qwen3.5, follows the release of Qwen3 on April 28, 2025, and Qwen 2.5 in September 2024. - The Mixture-of-Experts (MoE) architecture activates a subset of the model's total parameters for any given input, which significantly reduces computational cost and latency during inference. For instance, the Qwen3.5-397B-A17B model activates only 17 billion of its 397 billion parameters per forward pass. - Qwen3.5's predecessor, Qwen3, introduced a dual-mode reasoning feature that dynamically switches between a "thinking mode" for complex tasks like logical reasoning and code generation, and a "non-thinking mode" for quicker, general-purpose dialogue. - The Qwen family includes specialized models for different tasks, such as Qwen-VL for vision-language tasks (like analyzing images and documents), Qwen-Audio for audio processing, and Qwen2.5-Coder and Qwen2.5-Math for specialized coding and mathematics tasks. - Alibaba is positioning its Qwen models as a foundational layer for both consumer and enterprise applications, integrating them into its e-commerce platforms, productivity tools, and cloud services. The company is also exploring robotics and embodied AI through its Qwen lab. - The open-source Qwen models have gained significant traction, with the Qwen 2.5-72B-Instruct model topping the OpenCompass large language model leaderboard in September 2024, outperforming models like Claude 3.5 and GPT-4o in certain benchmarks. - For developers, the Qwen3 series supports a context length of up to 131,072 tokens and has demonstrated strong performance in code generation benchmarks. The models are available on platforms like Hugging Face and ModelScope under an Apache 2.0 license. - In the competitive landscape of open-source AI, Qwen 2.5 has been noted for its strong performance in mathematical reasoning and handling structured data, while competitors like Meta's Llama 3 have shown strengths in coding tasks and tokenization efficiency.
