Qwen Releases Efficient 35B Multimodal MoE Model

Alibaba's Qwen team has released Qwen3.5-35B-A3B, a new open-source model that brings a powerful combination of efficiency and multimodality. Available under a permissive Apache 2.0 license, the model continues the trend of advanced architectures being made available for commercial use and research.

The model's key feature is its Mixture of Experts (MoE) architecture. While it contains 35 billion total parameters, it only activates a fraction of them—around 3 billion—for any given inference task. This design allows it to access a vast store of knowledge while keeping computational costs significantly lower than a dense model of a similar size.

Beyond its text capabilities, Qwen3.5-35B-A3B is a Vision Language Model (VLM), meaning it can natively process and understand both text and images. This makes it suitable for a wide range of applications, from image captioning and visual question answering to more complex multimodal reasoning tasks.

This release provides developers with another strong option in the growing ecosystem of open MoE models. By combining an efficient architecture with multimodal understanding, Qwen3.5-35B-A3B offers a compelling balance of performance and practicality for building sophisticated AI applications.