Google Cloud backs Reflection's Beam for coding agents

Reflection has released Beam, its first open-weight model for software engineering and long-horizon agentic tasks. Google Cloud has added day-one support for deploying the model on its AI Hypercomputer and NVIDIA AI infrastructure.

The model is available under an Apache 2.0 licence, with BF16, FP8 and NVFP4 checkpoints at launch. Beam has about 501 billion total parameters, with around 23 billion active parameters per token, and an approximate context window of one million tokens.

Google Cloud worked with Reflection to tune Beam's serving stack using llm-d with vLLM and NVIDIA Dynamo with SGLang. The work focused on improving throughput and cutting latency for agentic coding workloads that call a model repeatedly across long sessions.

In a statement, Reflection linked the deployment work to measured gains in production-style use cases.

"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", said Joe Spisak, Vice President of Product and Head of Open Source at Reflection.

Beam is aimed at developers and researchers building coding agents and other systems that need repeated tool use over extended tasks. The model is designed for text-based software engineering, including resolving GitHub issues, running command-line and DevOps workflows, and completing longer tasks through coding harnesses already used by developers.

The launch adds another entrant to the growing market for open-weight models tuned for programming and reasoning. Reflection said Beam reaches solutions using three to four times fewer reasoning tokens than comparable open models, which it argued lowers the compute needed for each reasoning step.

Serving issues

Google Cloud's announcement focused less on model training and more on the challenge of running these systems at scale. In multi-turn coding sessions, a model may be called dozens of times while conversation history grows as file reads, code changes, compiler logs and tool outputs are added to the prompt.

That creates pressure at two points in inference. One is the prefill stage, where systems process a long prompt before generating an answer. The other is memory use for key-value caches, which store prior context so the model does not have to recompute the full history on every turn.

Google Cloud said standard stateless load balancing can send sequential turns from the same session to different replicas, forcing the system to recompute large prompts from scratch. That behaviour can raise latency and waste GPU resources as workloads scale.

The deployment recipes published alongside the launch use prefix-aware routing, prefill and decode separation, speculative decoding, and tiered cache storage across GPU memory, host DRAM and local solid-state drives. Google Cloud said these methods are intended to preserve Beam's token efficiency in production across multiple sessions.

Hardware options

Customers can run optimised Beam deployments on A4X Max instances based on NVIDIA GB300 NVL72 systems, and on G4 virtual machines using NVIDIA RTX PRO 6000 GPUs. Benchmarking showed up to five times higher throughput and eight times lower latency on A4X Max, and up to two times higher throughput and eight times lower latency on G4 virtual machines, compared with baseline deployments on the same hardware.

Google Cloud also noted that the two serving configurations were benchmarked on different hardware setups, so the results are intended as gains over separate baselines rather than a direct comparison between the two systems.

The software stack is being made available through Google Cloud's AI Hypercomputer recipes, the llm-d repository and Model Garden in the Gemini platform. That gives developers prebuilt deployment options instead of requiring them to assemble the serving architecture themselves.

Open model push

Beam's release under a permissive licence stands out at a time when access to large reasoning and coding models remains split between proprietary application programming interfaces and smaller open models. By publishing open weights, Reflection is targeting organisations that want more direct control over how models are run, adapted and deployed in their own environments.

For Google Cloud, day-one support is part of a broader effort to attract developers building AI agents that place sustained demands on infrastructure rather than one-off chatbot queries. Those workloads depend not only on model quality, but also on how efficiently cloud systems route requests, manage memory and maintain responsiveness as prompts grow longer over time.

Reflection said Beam works with tools including its Mirror command-line interface, as well as Claude Code, Hermes and OpenCode, and is intended to fit into existing coding workflows.