vLLM Supports Reflection Beam Model, Reports 2.71x Decode Latency Reduction

The vLLM project announced on October 16, 2026, that its inference engine now offers "Day 0" support for Beam-1.0, a new model released by Reflection. Beam is a Mixture-of-Experts (MoE) model with 501 billion total parameters and 23 billion active parameters per token.

According to the official vLLM blog post, the team utilized coding agents to autonomously optimize the model's execution on NVIDIA Blackwell hardware. These agents were responsible for profiling bottlenecks, writing kernels, and running benchmarks. vLLM reports that after seven key optimizations, batch-1 decode latency dropped from 8.95 ms to 3.30 ms per step under TP8 configuration, representing a 2.71x speedup. Prefill speed improved by 1.87x.

Additionally, vLLM integrated Beam's DFlash2 speculative decoding drafter and introduced adaptive verification mechanisms. Official data indicates a throughput of 708 tokens/s for a single request under TP8. In high-concurrency scenarios, combining Attention DP with MegaMoE technology yielded a 1.71x higher peak input throughput compared to traditional tensor parallelism.

It should be noted that these performance metrics are internal test results reported by vLLM and its partner Inferact, without independent third-party reproduction. The release primarily targets enterprise users and researchers with appropriate hardware resources.