Alibaba releases open weights for Qwen3.8-27B multimodal model

Alibaba’s Qwen team has released open weights for the Qwen3.8-27B, a 27-billion-parameter multimodal model built specifically for local deployment. The release represents a meaningful pivot for Alibaba, which had been keeping its flagship AI models behind proprietary walls. What Qwen3.8-27B actually is The Qwen3.8-27B handles text, image, and video inputs. Alibaba designed this model for practical local inference, targeting hardware with around 24GB of VRAM, the kind of GPU memory found in a high-end consumer graphics card like an NVIDIA RTX 4090. The 27B model sits at the lighter end of Alibaba’s new Qwen3.8 family. Its bigger sibling, the Qwen3.8-Max, features 2.4 trillion total parameters with roughly 95 billion active at any given time. That model requires multi-node infrastructure and is aimed at enterprise-scale data center deployments. Open weights for the Max version became available around mid-August under the label Qwen3.8-2.4T-A95B. Why Alibaba is going open Prior to the Qwen3.8 announcement, several Max-class models were maintained under proprietary access. The shift toward open weights for both the 27B and the Max marks a return to the open-sourcing approach that chara...