NVIDIA Optimizes Alibaba Qwen3.8-Max on GB300 NVL72

Alibaba has launched the open weights for Qwen3.8-2.4T-A95B, also known as Qwen3.8-Max. This massive model features 2.4 trillion total parameters, though it only activates 95 billion parameters per token. Built as a fine-grained mixture of experts model, it utilizes a hybrid of full and linear attention. It supports a context window of up to one million tokens and can generate outputs up to 128,000 tokens. This architecture is specifically designed to handle demanding agentic workloads, such as multi-step workflows and large-scale document analysis.

To run a model of this scale, NVIDIA has optimized the system for its GB300 NVL72 platform. This rack-scale system links 72 Blackwell Ultra GPUs together with a 130 terabytes-per-second NVLink domain, which eliminates communication bottlenecks. Running the model in FP8 precision, the setup achieves a throughput of more than 4,000 tokens per second per GPU and over 350 tokens per second per user. NVIDIA plans to introduce further optimizations in the future, including NVFP4 precision, to boost these speeds even more.

For practitioners, this release makes frontier-level AI capabilities highly accessible and configurable. The model features built-in reasoning controls that allow developers to select low, high, or extra-high settings. This lets them trade compute power for reasoning depth depending on whether they are doing complex coding or high-throughput document processing. Furthermore, developers can deploy the model using open-source stacks like SGLang, vLLM, and NVIDIA Dynamo, or package it via a model-free NVIDIA NIM container. For customization, NVIDIA NeMo AutoModel supports supervised fine-tuning and low-rank adaptation directly on Hugging Face checkpoints.