AWS has published deployment guidance for Reflection AI’s Beam, an open-weight model designed for agentic coding and long-horizon tasks. According to the AWS post, Beam is intended to maintain and modify code, use tools, interpret intermediate results, and complete multi-step workflows.

Reflection AI says Beam is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active parameters. The company reports an effective context length of 1 million tokens. Its weights are available through Hugging Face under the Apache 2.0 license, with Beam 1.0 checkpoint variants in BF16, FP8, and NVFP4 formats.

AWS presents two main deployment paths. SageMaker JumpStart provides a managed real-time endpoint, while SageMaker HyperPod Inference offers Kubernetes-based control over distributed GPU infrastructure. HyperPod deployments can use the AWS-managed JumpStart container or custom SGLang and vLLM runtimes. The post says these options allow users to manage aspects such as serving configuration, distributed execution, caching, routing, and scaling.

AWS states that Beam requires a multi-GPU deployment because of its total parameter count. The guidance identifies H100 and H200 instances for FP8 configurations and B200 instances for NVFP4 configurations, while noting that users should confirm checkpoint, runtime, and parallelism support for their selected setup. At launch, AWS says Beam was available through SageMaker JumpStart in the US East (N. Virginia) and US East (Ohio) regions.