Now in Foundry: Qwen3.5 Medium Model Series

The Qwen3.5 Medium Model Series, available in Microsoft Foundry, is a set of Vision Language Models (VLMs). These models utilize early-fusion multimodal training and support 201 languages within a 262K context window. There are three models: a 27B dense model, and two Mixture-of-Experts (MoE) models of 35B and 122B sizes. All models were trained using unified vision-language training, gated delta networks, and scalable RL. The 27B model is optimized for latency-sensitive applications, demonstrating strong instruction following capabilities. The 35B-A3B model excels in throughput, utilizing MoE for cost-effective inference. The 122B-A10B model showcases the highest performance and is suited for complex multimodal tasks. Each model has suggestions of use cases, from circuit board quality control to contract review and financial research. These models can be deployed through Microsoft Foundry, utilizing the Hugging Face model catalog integration.