Reflection released Beam, its first open-weight model aimed at frontier software engineering and long-horizon agentic tasks, and Google Cloud announced day-0 support with ready-to-run deployment recipes on AI Hypercomputer and NVIDIA AI Infrastructure. Beam is released under Apache 2.0 with BF16, FP8, and NVFP4 checkpoints available at launch. It has roughly 501B total parameters with about 23B active per token and an approximately 1M-token context window. According to the post, Beam solves complex reasoning and coding problems using 3x to 4x fewer reasoning tokens than comparable open-weight models, and the optimized serving stack with llm-d (with vLLM) and NVIDIA Dynamo (with SGLang) delivers up to 5x higher performance and 8x lower latency on A4X Max instances (NVIDIA GB300 NVL72), with up to 2x higher throughput and 8x lower latency on G4 VMs (NVIDIA RTX PRO 6000), based on customer benchmarking.

Beam targets text-native software engineering and multi-step tool use, and plugs into coding harnesses like Reflection‘s Mirror CLI, Claude Code, Hermes, and OpenCode. But the post argues that inference efficiency in agentic workloads depends on both model-level reasoning efficiency and system-level serving efficiency. In a real agentic loop, the model is called dozens of times per session and the coding harness appends file reads, diffs, compiler logs, and tool outputs to a shared history that grows to tens or hundreds of thousands of tokens. Standard stateless Kubernetes round-robin routing sends later turns to replicas that must recompute the entire prompt, wasting GPU compute and creating prefill queues. Long-context KV caches also fill GPU HBM, and evicted blocks cause cache hit rates to drop and throughput to stall.

Google Cloud’s architecture on AI Hypercomputer addresses these bottlenecks with a vertically integrated stack. Prefix- and load-aware routing (GKE plus llm-d EPP or Dynamo router) tracks cached prefixes and queue depth, steering each follow-up turn to the replica holding its warm KV prefix while applying turn-priority flow control. Disaggregated prefill and decode pools ensure heavy prompt bursts do not stall token generation. Native Blackwell NVFP4, Expert Parallelism for Beam‘s 23B active experts, Data-Parallel Attention, and DFlash2 speculative decoding are used across A4X Max and G4 instances. A tiered KV cache keeps active blocks in GPU memory (Tier 1) and offloads warm evicted blocks to host DRAM and local NVMe SSDs (Tier 2) via vLLM’s OffloadingConnector or SGLang’s HiCache, letting historical context be fetched in milliseconds instead of recomputed.

The post includes benchmark notes that the two optimized configurations are not comparable to each other: NVIDIA Dynamo ran on A4X Max instances with 32 GB300 NVL72 GPUs, while llm-d ran on G4 VMs with 8 RTX PRO 6000 GPUs. Each chart shows gain over a baseline on the same hardware, and the uplift grows as more replicas give the router more warm KV caches and more room for prefill/decode separation. Layering optimizations shifts the Pareto frontier, improving output throughput and per-user interactivity together. All Kubernetes manifests, router configurations, and benchmarking scripts are open-source. Day-0 guides and recipes are available in AI-Hypercomputer/gpu-recipes, the llm-d repo, and Model Garden in Gemini platform. Weights are on Reflection‘s Hugging Face page, and the team will demo the stack at PyTorch Conference North America in San Jose and NVIDIA GTC Berlin.