Google Cloud 为 Reflection 的 Beam 模型提供 Day 0 支持：借助 llm-d 与 NVIDIA Dynamo 实现 5 倍性能提升

Today, Reflection released Beam, its first open-weight model built to give developers, researchers, and enterprises open control and customization for frontier software engineering and long-horizon agentic tasks. Released under a permissive Apache 2.0 license with BF16, FP8, and NVFP4 checkpoints available at launch, Beam sets a new standard for inference efficiency in the 500B-parameter class by solving complex reasoning and coding problems with 3x to 4x higher reasoning compute efficiency than comparable open models. Alongside today’s release, Google Cloud is announcing day 0 support and ready-to-run deployment recipes for serving Beam on Google Cloud AI Hypercomputer and NVIDIA AI Infrastructure. To help developers get the full benefit of Beam’s token efficiency in production, we worked with Reflection to optimize performance of the underlying serving stack using llm-d (with vLLM) and NVIDIA Dynamo (with SGLang) on Google Cloud: "By deploying our Beam model on Google Cloud, we achieved up to 5x higher performance and 8x lower latency to optimize serving of agentic workloads." - Joe Spisak, VP of Product & Head of Open Source, Reflection With today’s day 0 release on Google Cloud, developers get: Compounding model and system efficiency: Combining Beam’s 3x to 4x reasoning compute efficiency with system-level optimizations on A4X Max instances (NVIDIA GB300 NVL72) for up to 5x higher throughput and 8x lower latency,* and G4 VMs (NVIDIA RTX PRO 6000) for up to 2x higher throughput and 8x lower latency.* High-efficiency multi-turn serving: KV-cache-aware routing, prefill/decode disaggregation, speculative decoding and hierarchical KV cache offloading to Host DRAM and Local SSD reduce redundant prompt recomputation across long coding sessions. Turnkey optimized deployment: Our recommended performance optimized recipes for agentic workloads with llm-d and NVIDIA Dynamo are available in AI-Hypercomputer/gpu-recipes, llm-d repo and Model Garden in Gemini platform with a few clicks. *Based on customer benchmarking Scaling inference requires system-level optimization When evaluating models for agentic coding and tool use, inference efficiency happens at two distinct layers: how efficiently the model reasons and how efficiently the serving system executes multi-turn loops. Model-level efficiency: Reflection’s Beam Reflection designed Beam (~501B total parameters with ~23B active parameters per token, and an approximately 1M-token context window) specifically for text-native software engineering and multi-step tool use. It plugs directly into the coding harnesses developers already use, including Reflection’s Mirror CLI, Claude Code, Hermes, OpenCode, and , to resolve GitHub issues, run CLI and DevOps workflows, and complete long-horizon tasks. Because Beam activates only 23B parameters per token and reaches solutions using 3x to 4x fewer reasoning tokens than other comparable open weight models, it dramatically reduces the inference-time compute required for each reasoning step. System-level efficiency: AI Hypercomputer optimized serving Once developers deploy an agent into production, a second set of bottlenecks could appear at the serving layer. In an agentic coding loop, the model is called dozens of times in a single session. With every turn, the coding harness appends new file reads, code diffs, compiler logs, and tool outputs to a shared conversation history that quickly grows to tens or hundreds of thousands of tokens. Even when the model itself generates 3x to 4x fewer reasoning tokens on each turn, the accumulated context window across turns creates two major infrastructure challenges: First, prefill compute becomes a bottleneck. Standard Kubernetes load balancers distribute requests using stateless round-robin routing. When a coding harness sends Turn 15 of a session, it often lands on a different replica than Turn 14. That replica is forced to recompute the entire historical prompt from scratch, wasting GPU compute and creating long prefill queues that delay other users. Second, decode becomes bottlenecked by KV cache capacity. Holding long-context KV caches for dozens of concurrent agent sessions quickly fills up GPU High-Bandwidth Memory (HBM). Once HBM fills up, the serving engine has to evict older KV blocks. As soon as those blocks are evicted, cache hit rates drop and GPU throughput stalls. To preserve Beam’s model-level efficiency at cluster scale, the serving stack has to be optimized as a distributed system. That is why our recipes pair Beam with llm-d (with vLLM) and NVIDIA Dynamo (with SGLang). Inside the AI Hypercomputer inference architecture To run Beam’s BF16, FP8, and NVFP4 checkpoints at peak efficiency on Day 0, we designed a vertically integrated architecture on AI Hypercomputer where each layer handles a specific part of the agentic serving lifecycle. System optimized architecture for serving agentic traffic Rather than treating inference pods as stateless endpoints, each layer of the stack is co-designed around the multi-turn agent lifecycle: Prefix- and load-aware routing (GKE + llm-d EPP / Dynamo router): Tracks cached prefixes and queue depth across worker pods, steering each follow-up turn from a coding harness to the replica holding its warm KV prefix while preventing queue saturation via turn-priority flow control. Disaggregated compute and speculative decoding (A4X Max & G4 instance): Runs vLLM and SGLang with native Blackwell NVFP4, Expert Parallelism (EP) for Beam’s 23B active experts, Data-Parallel Attention, and DFlash2 speculative decoding. Separating Prefill and Decode (P/D) pools ensures heavy prompt bursts never stall active token generation. Tiered KV cache (GPU Memory -> DRAM & Local SSD) Active KV blocks stay in high-bandwidth GPU memory (Tier 1), while warm KV blocks evicted during long-horizon sessions overflow into host CPU DRAM and local NVMe SSDs (Tier 2) via vLLM’s OffloadingConnector or SGLang's HiCache, allowing historical context to be fetched in milliseconds instead of recomputed from scratch. System optimizations for multi-turn agentic workloads When serving multi-turn coding and tool-use workloads like the AgentX trajectory dataset across rising concurrency levels, each layer in the serving stack contributes to the performance and latency improvements for multi-turn, long context traffic patterns of agentic workloads: A note on reading these charts. The two configurations are not comparable to each other. NVIDIA Dynamo ran on A4X Max instances with 32 x NVIDIA GB300 NVL72GPUs; llm-d ran on G4 VMs with 8 x NVIDIA RTX PRO 6000 GPUs. Each chart shows the gain over a baseline on the same hardware. More replicas give the router more warm KV caches to hit and more room to split prefill from decode, so the uplift grows with the deployment. Both llm-d and NVIDIA Dynamo scale up and down to match the underlying infrastructure. NVIDIA Dynamo optimized perf vs baseline deployment of Reflection's Beam llm-d optimized perf vs baseline deployment of Reflection's Beam Across both platforms, layering these optimizations shifts the Pareto frontier to deliver higher output throughput and faster per-user interactivity simultaneously: Baseline deployment (Stateless round-robin routing). In a standard Kubernetes deployment without prefix awareness, consecutive turns from the same coding session land on different replicas, forcing constant prefill recomputation on historical context and causing throughput to plateau as concurrency scales. KV-cache-aware & flow control routing. Enabling prefix-aware routing in NVIDIA Dynamo or llm-d’s Sticky EPP with turn-priority flow control keeps multi-turn agent trajectories pinned to the replica holding their warm KV cache, immediately cutting Time-to-First-Token (TTFT) latency and freeing up GPU compute, resulting in a significant increase in token throughput. Host DRAM and Local SSD (LSSD) KV offloading. Offloading evicted KV blocks from GPU memory into host DRAM and local NVMe SSDs prevents cache thrashing when dozens of long-context coding sessions exceed on-device GPU memory capacity. Disaggregated serving, speculative decoding, and KV pool tuning. Tuning hybrid KV cache memory allocation, enabling speculative decoding, and disaggregating prefill and decode workers isolates token generation from heavy prompt bursts while the draft model multiplies output throughput to keep per-agent token interactivity fast. Get started and meet us at PyTorchCon and GTC Berlin This blog explored how you can leverage system level optimizations to scale performance to serve Reflection's Beam family of models for agentic workloads. All of the Kubernetes manifests, router configurations, and benchmarking scripts from this post are open-source and ready to run on your own GKE cluster. Deploy the optimized stack: Explore the day 0 guides for running Reflection’s Beam with llm-d and NVIDIA Dynamo in the AI-Hypercomputer/gpu-recipes repository on A4X Max and G4, and Model Garden in Gemini platform. Download the weights: Grab the Apache 2.0 BF16, FP8, and NVFP4 checkpoints on Reflection’s Hugging Face page. Visit us at the llm-d and PyTorch Conference in San Jose: If you are attending PyTorch Conference North America in San Jose this October, stop by the llm-d community sessions and the Google Cloud booth #. We will be walking through live demos of KV-cache-aware routing, P/D disaggregation, and KV cache tiering for agentic workloads. Come see us at NVIDIA GTC Berlin: If you are attending NVIDIA GTC Berlin, come connect with us at our booth #3026, or our Google Cloud Hub at w3.hub: just a short 5 minute walk from the venue. Learn more about our sessions.