vLLM adds day-0 serving for Reflection's Beam 501B MoE model
Original titleServing Reflection Beam on vLLM from Day 0
AISummary
vLLM supports Reflection's Beam, a 501B-parameter MoE with 23B active parameters per token, from day 0 and optimizes it on NVIDIA Blackwell. Agent-driven kernel work cut batch-1 decode from 8.95 ms to 3.30 ms per step on TP8. A DFlash2 drafter with adaptive verification reaches 708 tokens/s on a single request and adds 10.8% throughput at concurrency 128.
AIWhy it matters
The post names the specific kernel fusions, parallelism setups, and measured latency and throughput numbers that let engineers judge what serving this model on vLLM involves.
Source: vLLM Blog · vllm.aiPublished · added here