vLLM announced Day 0 support for Qwen3.8-2.4T-A95B, a 2.4-trillion-parameter sparse mixture-of-experts model with 512 experts, running without architecture modifications. The team built optimized kernels for NVIDIA and AMD hardware to handle the model’s linear attention layers and MoE routing, while FP4-quantized variants preserved accuracy, scoring 90.37% on GSM8K versus 89.61% for FP8. Full-precision deployment requires at least two GPU nodes, while quantized weights fit on a single node, with recommended generation parameters and client code provided for inference.
