Serve Qwen3.8-2.4T-A95B, a 2.4T-Parameter Model, with Configurable Reasoning on NVIDIA GB300 NVL72

Alibaba and NVIDIA Partner to Serve 2.4T Parameter Open-Weight Model

Alibaba has released the open weights for Qwen3.8-2.4T-A95B, its largest open-source model, making near-frontier AI capabilities accessible to the broader developer ecosystem. The release is accompanied by a deep collaboration with NVIDIA to optimize the model for its data-center-scale hardware. This partnership is critical for deploying a model of this magnitude, ensuring that its advanced capabilities for reasoning and agentic workloads are practical for production use on platforms like the NVIDIA GB300 NVL72.

Technical Specifications and Architectural Innovations

The Qwen3.8-2.4T-A95B model is engineered for complex, long-running tasks through several architectural innovations. Its fine-grained Mixture-of-Experts (MoE) design makes the 2.4 trillion total parameter count manageable by activating only 95 billion parameters per token, tying serving costs to active compute rather than the full model size. A hybrid attention mechanism, which alternates between full and linear attention, allows the model to handle a massive one-million-token context window without unbounded growth in compute and memory requirements. Initial benchmarks on the NVIDIA GB300 NVL72 show a throughput of over 4,000 tokens per second per GPU in FP8 precision.

• Total Parameters: 2.4T
• Active Parameters: 95B per token
• Architecture: Fine-grained MoE with hybrid full/linear attention
• Context Window: Up to 1 million tokens
• Output Length: Up to 128K tokens
• Reasoning Control: Built-in controls (low/high/xhigh) to trade compute for quality

The availability of this model on day-zero with support from NVIDIA's inference stacks, including SGLang, vLLM, and the NVIDIA NIM microservice, streamlines the path to production for developers. The collaboration extends to post-training, with NVIDIA NeMo AutoModel enabling users to fine-tune the model for domain-specific applications. This integrated approach highlights a growing trend where the release of a state-of-the-art open model is deeply intertwined with the software and hardware ecosystem required to run it effectively.

The joint release of Qwen3.8-2.4T-A95B and its optimized serving recipes for the GB300 NVL72 solidifies a key market dynamic: the practical value of massive open-source models is now inseparable from co-optimization with specific, vertically integrated hardware platforms. This makes large model releases a powerful ecosystem-building tool for hardware providers, anchoring developers to their high-performance computing stack.