The recipes target Google Cloud AI Hypercomputer and NVIDIA AI Infrastructure, and are available through AI-Hypercomputer/gpu-recipes, the llm-d repository, and Model Garden in the Gemini platform.
Beam is released under Apache 2.0 with BF16, FP8, and NVFP4 checkpoints. Google describes it as a roughly 501B-parameter model with about 23B active parameters per token and an approximately 1M-token context window. The company says Beam can be used with coding harnesses including Mirror CLI, Claude Code, Hermes, and OpenCode.
Google’s serving configurations pair llm-d with vLLM or NVIDIA Dynamo with SGLang. It reports up to 5x higher throughput and 8x lower latency on A4X Max instances, and up to 2x throughput with 8x lower latency on G4 VMs, each against a baseline on the same hardware. These are customer-benchmark claims; the source does not establish independent results, general availability terms, pricing, or performance on other workloads.