Qwen releases flagship 397B multimodal MoE

The Qwen team at Alibaba has released Qwen3.5-397B-A17B, a powerful new open-source model that pushes the boundaries of scale and efficiency. As detailed on its Hugging Face repository, the model features a staggering 397 billion total parameters, making it one of the largest open models available.

A Sparse Architecture at Scale

What makes this scale manageable is its Mixture-of-Experts (MoE) architecture. Instead of activating all 397 billion parameters for every task, the model intelligently routes queries through a smaller subset, using only 17 billion active parameters at inference time. This "sparse" approach allows for the vast knowledge capacity of a huge model while keeping computational demands relatively low.

Beyond its scale, Qwen3.5 is also a capable vision-language model (VLM). This multimodal capability means it can process and understand both text and images, enabling more complex applications in areas like image captioning, visual question answering, and content analysis.

Released under the permissive Apache 2.0 license, Qwen3.5-397B-A17B represents a significant contribution to the open-source AI ecosystem. By providing access to a flagship-class MoE model, Alibaba is enabling developers and researchers to build on top of state-of-the-art multimodal AI without the need to train such a massive model from scratch.