Google Cloud adds optimized deployment recipes for Reflection's Beam

Google Cloud released deployment recipes for Reflection's Beam coding model, adding support through its AI Hypercomputer infrastructure and Model Garden. Google says customer benchmarks showed up to 5x higher throughput and 8x lower latency on A4X Max instances against a baseline deployment on the same hardware. The recipes use llm-d with vLLM or NVIDIA Dynamo with SGLang. They route follow-up requests to servers holding reusable conversation data, separate prompt processing from token generation, and move cached data into host memory and local storage when GPU memory fills. The performance gains concern multi-turn agent workloads. Google cautions that the two configurations use different hardware and cannot be compared directly.