Open-Source Multimodal AI Model 'Qwen3.5' Released
An open-weight, 397-billion-parameter multimodal AI model named Qwen3.5 has been released. The model uses a Mixture-of-Experts (MoE) architecture for efficiency, supports over 200 languages, and is designed as a native multimodal agent for images, text, and other data types. Its accessibility is intended to lower the barrier for developers to experiment with advanced AI on edge devices.
- The Qwen series of models is developed by Alibaba Cloud and the name approximately translates to "thousand questions with general meaning". The project includes a family of models of various sizes and specializations, such as Qwen-VL for vision-language tasks and Qwen-Coder. - The Mixture-of-Experts (MoE) architecture activates only a fraction of the model's total parameters for any given input, making it more efficient. For Qwen3.5, only 17 billion of its 397 billion parameters are active per token, which significantly increases speed and reduces operational costs. This allows it to be 60% cheaper to run and have eight times higher throughput for large workloads compared to its predecessor, Qwen3-Max. - Qwen3.5 integrates native multimodal capabilities from the pre-training stage, allowing it to process text, images (up to 1344x1344 resolution), and video clips within a single system. This "early fusion" architecture is a key design feature. - Alibaba claims that on 80% of evaluated benchmarks, Qwen3.5 outperforms models like GPT-5.2, Anthropic's Claude Opus 4.5, and Google's Gemini 3 Pro. Specifically, it scores 90.8 on OmniDocBench v1.5 for document recognition, surpassing its competitors. - A closed-source, API-only version called Qwen3.5-Plus is also available through Alibaba's Model Studio. This version offers a significantly larger context window of 1 million tokens, compared to the open-weight model's 256,000 tokens. - The open-weight version of the model is licensed under Apache 2.0 and can be deployed locally using frameworks such as Ollama, LM Studio, and vLLM. The model weights are available for download from platforms like Hugging Face and ModelScope. - This release is part of a broader trend of rapid development in China's AI sector, with competitors like ByteDance's Doubao and DeepSeek also releasing new models. The Qwen family of models has seen significant adoption, with downloads on Hugging Face surpassing 40 million.
