Google Cloud introduced launch-day Beam deployment recipes using llm-d and NVIDIA Dynamo, including support for NVIDIA GB300 NVL72 systems.

Google Cloud announced launch-day support for Reflection’s Beam on Google Cloud AI Hypercomputer and NVIDIA AI Infrastructure. Its optimized serving recipes use llm-d with vLLM and NVIDIA Dynamo with SGLang, following collaboration with Reflection.

Google reports up to 5 times higher throughput and 8 times lower latency on A4X Max instances using NVIDIA GB300 NVL72 systems. For G4 virtual machines using NVIDIA RTX PRO 6000 GPUs, it reports up to 2 times higher throughput and 8 times lower latency; these are customer-benchmarking claims, and the supplied excerpt does not specify the full comparison baseline or test configuration.

The serving techniques include key-value (KV) cache-aware routing, separation of prefill and decode processing, speculative decoding, and hierarchical KV-cache offloading to host memory and local solid-state drives. Google states that these techniques reduce redundant prompt recomputation across long coding sessions.

The recommended deployment recipes are available through AI-Hypercomputer/gpu-recipes, the llm-d repository, and Model Garden in the Gemini platform. Google distinguishes these system-level serving results from Reflection’s separate claims about Beam’s reasoning-token efficiency.