Serve Qwen3 2.4T-Parameter Model with Configurable Reasoning on Nvidia GB300 (developer.nvidia.com)

Alibaba has unveiled the open weights for Qwen3.8-2.4T-A95B, its most advanced open-weight AI model, featuring an impressive 2.4 trillion parameters with 95 billion activated per token. This model employs a fine-grained mixture of experts (MoE) architecture, combining full and linear attention with the ability to handle a context window of up to one million tokens and an output length of 128,000. Its deployment requires sophisticated, data-center-scale compute resources, and NVIDIA is optimizing kernels and software to facilitate multinode deployments. At launch, the model boasts a throughput exceeding 4,000 tokens per second per GPU, establishing its capability for intricate reasoning tasks such as coding and large-scale document analysis. The significance of Qwen3.8 lies in its ability to maintain high performance while managing extensive context, thanks to its hybrid architecture that intelligently switches between full and linear attention layers. This approach helps optimize compute and memory usage as context scales, crucial for agentic applications that require ongoing instruction and memory tracing. The model's innovative MoE methodology allows for a more efficient use of resources by activating only the necessary experts for each token, significantly reducing operational costs compared to dense models. Developers can further customize performance through built-in reasoning controls and leverage NVIDIA’s rich set of tools for domain-specific tuning, positioning Qwen3.8 as a groundbreaking asset for AI development and production.