Deploying Qwen3.8-2.4T-A95B on Amazon SageMaker HyperPod
Alibaba's open-weight Qwen3.8-2.4T-A95B model is successfully deployed on infrastructures featuring NVIDIA B300 GPUs using Amazon SageMaker HyperPod and vLLM.
The Qwen3.8-2.4T-A95B model, developed by Alibaba, has been released with a technical guide for running on Amazon SageMaker HyperPod using vLLM and NVIDIA B300 Blackwell Ultra GPU infrastructure.
Prominent Architectural Features of the Qwen3.8 Model
Published by the Alibaba Qwen team on August 12, 2026, the model represents the first time a Qwen-Max class model has been made available with open weights.
Boasting a total of 2.4 trillion parameters, the model utilizes 95 billion active parameters per token and supports a 262K token context window with its hybrid architecture.
Infrastructure and Hardware Requirements
According to the technical details provided, the deployment process is executed on ml.p6-b300.48xlarge instances featuring eight NVIDIA B300 Blackwell Ultra GPUs.
During the setup phase, cluster provisioning and NVFP4 quantization techniques are used to compress the model size to approximately 1.2 terabytes so it fits on a single node.
Advanced vLLM Integration and Capabilities
The configuration process encompasses built-in reasoning controls, tool-calling features, and native Multi-Token Prediction speculative decoding mechanisms.