Qwen Releases 35B Multimodal Mixture-of-Experts Model

Alibaba's Qwen team has released a new large multimodal model, Qwen3.6-35B-A3B, adding a sophisticated new option to the open-source AI landscape. As a Vision Language Model (VLM), it is designed to understand and reason about both text and images, making it suitable for a wide range of complex tasks.

The model's key feature is its Mixture-of-Experts (MoE) architecture. While it contains a total of 35 billion parameters, it only activates approximately 3 billion of them for any given input. This sparse activation approach allows the model to achieve the performance of a much larger dense model while requiring significantly less computational power during inference, a crucial advantage for practical deployment.

Why it matters

Qwen3.6-35B-A3B represents a significant contribution to the growing field of efficient, high-performance open models. By combining a multimodal design with a sparse MoE architecture, it provides researchers and developers with a powerful tool that balances capability with computational cost.

The model has been released on Hugging Face under a permissive Apache 2.0 license, encouraging broad adoption and further innovation within the community. Its release continues the trend of major AI labs making increasingly powerful foundation models accessible to the public.