Qwen Releases Flagship 122B Multimodal MoE Model

Alibaba's Qwen team has unveiled Qwen3.5-122B-A10B, the flagship model in its new Qwen3.5 series. This release marks a significant entry into the high-end open-source model space, targeting complex multimodal tasks with a powerful and efficient architecture.

The model's main innovation is its scale combined with its structure. It is a 122-billion-parameter Mixture-of-Experts (MoE) model, but only activates 10 billion parameters for any given input. This design provides the knowledge capacity of a very large model while keeping the computational cost for inference relatively low, making it more accessible than a dense model of a similar size.

A Vision-Language Specialist

As a vision-language model (VLM), Qwen3.5-122B-A10B is designed to understand and process information from both images and text simultaneously. This makes it suitable for a wide range of applications, including visual question answering, image captioning, and object recognition within a broader textual context.

In keeping with the project's commitment to open-source development, the model is available under the permissive Apache 2.0 license, which allows for both academic and commercial use. Developers can access the model weights and explore its capabilities on the Hugging Face Hub.