Open-Source Multimodal AI Model Qwen3.5 Released
A new 397-billion-parameter open-weight multimodal model named Qwen3.5 has been released. The model, which supports over 200 languages, was built using a Mixture-of-Experts (MoE) architecture. It is specifically designed for efficient inference, making it suitable for deployment on edge and embedded systems, thereby lowering the barrier for developers to experiment with state-of-the-art AI.
- While the total parameter count is 397 billion, its Mixture-of-Experts (MoE) architecture activates only 17 billion parameters for any given token, a design that significantly reduces computational cost. - The model's architecture combines Gated Delta Networks, a form of linear attention, with the MoE structure, enabling it to process long data sequences with much greater efficiency. This results in decoding throughput that is between 8.6 and 19 times faster than its predecessor, Qwen3-Max. - Qwen3.5 was designed with "early fusion" training, meaning it learned from multimodal data (text, images, UI screenshots) from the beginning, rather than having vision capabilities added to a text-only base. - It demonstrates strong performance on various benchmarks, scoring 88.4 on the graduate-level reasoning test GPQA Diamond and 76.5 on the instruction-following evaluation IFBench. In agentic search tasks, it outperformed Gemini 3 Pro. - The model is released under the Apache 2.0 license, which permits commercial use. A hosted version with a 1-million token context window, called Qwen3.5-Plus, is available through Alibaba Cloud. - Due to extensive training on user interface screenshots, the model can perform "visual agent" tasks like navigating apps, filling out forms, and organizing files based on natural language commands. - The open-weight model has a native context length of 262,144 tokens and can be run on a server with eight H100 GPUs using inference engines like vLLM and SGLang. - Qwen is a large language model family developed by Alibaba Cloud, with the first models released in April 2023. The open-source versions of the Qwen series have been downloaded millions of times.
