Key Points
- Google Cloud announces day 0 support for Reflection's Beam open-weight AI model
- Deployment on Google Cloud infrastructure claims up to 5x higher throughput
- Beam uses 23 billion active parameters per token from 501 billion total
Google Cloud has announced immediate deployment support for Reflection’s Beam, a 501 billion parameter open-weight artificial intelligence model designed for software engineering and automated task execution, the company said.
The model, released under an Apache 2.0 licence, becomes available on Google Cloud AI Hypercomputer and NVIDIA AI Infrastructure with what the company describes as optimised serving configurations. Beam activates only 23 billion parameters per token, a technique called mixture of experts that reduces the computational load during inference, the process where a trained model generates outputs from new inputs.
According to Google Cloud, deploying Beam on its infrastructure with llm-d and NVIDIA Dynamo optimisation frameworks achieves up to five times higher throughput and eight times lower latency compared to baseline deployments. These figures are based on customer benchmarking, the company said, though it did not name the customers involved.
Joe Spisak, vice president of product and head of open source, Reflection, said the deployment delivered significant performance improvements for agentic workloads, a category of AI applications where models execute multi-step tasks autonomously rather than responding to single queries.
Token efficiency
Beam is designed to solve complex reasoning and coding problems using three to four times fewer reasoning tokens than comparable open models, according to Reflection. Tokens are the units of text that language models process, and reducing token consumption directly lowers the computational cost of running the model.
The model includes an approximate one million token context window, meaning it can process and retain information from roughly one million tokens of input text. This capacity is relevant for coding applications where the model must track long files, multiple code changes and extended conversation histories.
Reflection designed Beam to integrate with existing developer tools including its own Mirror CLI, Claude Code, Hermes and OpenCode. The model is intended to resolve GitHub issues, run command line and DevOps workflows and complete tasks that span multiple steps, the company said.
Google Cloud’s deployment recipes address what it describes as system-level bottlenecks that emerge when AI models are used repeatedly in coding sessions. In these scenarios, each new query from a developer appends additional context, file changes and tool outputs to a growing conversation history that can reach hundreds of thousands of tokens.
Infrastructure architecture
The serving architecture uses several techniques to manage these extended sessions. KV cache aware routing, where KV cache refers to the stored results of previous computations that can be reused rather than recalculated, directs follow-up requests to the same server replica holding the relevant cached data. This avoids forcing a new server to recompute the entire conversation history from scratch.
By the numbers
Key figures from this story- 501 billion
- total parameters in Beam model
- 5x
- claimed throughput improvement on A4X Max
- 23 billion
- active parameters per token
The system separates prefill and decode operations into distinct pools. Prefill is the initial processing of a user’s input, while decode is the generation of output tokens. Separating these prevents heavy input processing from delaying active token generation for other users.
When GPU memory fills with cached data from multiple concurrent sessions, the system offloads older cache blocks to host CPU memory and local solid state drives. This tiered storage approach allows historical context to be retrieved in milliseconds rather than recomputed, according to Google Cloud.
The deployment supports Beam’s BF16, FP8 and NVFP4 checkpoints, different numerical precision formats that trade off between model accuracy and computational efficiency. Lower precision formats like NVFP4 require less memory and compute but may reduce output quality in some applications.
Google Cloud tested the optimised deployments on two hardware configurations. A4X Max instances using NVIDIA GB300 NVL72 GPUs achieved up to five times higher throughput and eight times lower latency. G4 virtual machines using NVIDIA RTX PRO 6000 GPUs achieved up to two times higher throughput and eight times lower latency.
The company noted these configurations are not directly comparable as they use different hardware. The performance gains in each case are measured against a baseline deployment on the same hardware without the optimisations.
Speculative decoding, a technique where a smaller draft model proposes multiple tokens that the main model then verifies in parallel, contributes to the throughput improvements. This approach multiplies output speed when the draft model’s predictions are accurate.
The deployment recipes are available through Google Cloud’s AI Hypercomputer GPU recipes repository, the llm-d repository and Model Garden in the Gemini platform. Google Cloud described these as turnkey configurations that can be deployed with minimal setup.
Reflection released Beam with checkpoints in three precision formats to give developers options for balancing performance against resource requirements. The Apache 2.0 licence allows commercial use and modification of the model weights.
Your Questions, Answered
What is Reflection's Beam model?
Beam is a 501 billion parameter open-weight AI model released under Apache 2.0 licence, designed for software engineering and automated multi-step tasks. It activates only 23 billion parameters per token to reduce computational requirements.
What performance improvements does Google Cloud claim?
Google Cloud claims up to five times higher throughput and eight times lower latency on A4X Max instances, and up to two times higher throughput on G4 virtual machines, compared to baseline deployments.
What is KV cache aware routing?
KV cache aware routing directs follow-up requests in a coding session to the same server holding cached computation results, avoiding the need to recompute entire conversation histories from scratch.
Which developer tools does Beam support?
Beam integrates with existing tools including Reflection's Mirror CLI, Claude Code, Hermes and OpenCode for resolving GitHub issues and running DevOps workflows.
