Qwen3.8 brings a 2.4T model to vLLM

Qwen3.8-2.4T-A95B is now supported in vLLM on day one, with 2.4 trillion total parameters and FP4 variants that let the model run on a single node instead of the two NVIDIA B300 or AMD MI355X nodes required by the larger-precision deployment. The vLLM release note says the NVFP4 and MXFP4 checkpoints preserve the team’s initial quality checks while the stack is validated across NVIDIA and AMD hardware. The operator fact to file beside Qwen3.8-Max’s nine-H200 preview burden is that quantization and serving support—not parameter count alone—now decide whether an open frontier model is actually deployable.