Introducing Reflection Beam: an open-weight model for agentic coding on AWS
aws ·
Reflection AI has introduced Reflection Beam, an open-weight model designed for agentic coding and long-horizon tasks such as code maintenance and multi-step workflows. The model features a sparse mixture-of-experts architecture with 501 billion total parameters and 23 billion active parameters, supporting an effective context length of one million tokens. Released under the Apache 2.0 license, Beam is available in BF16, FP8, and NVFP4 checkpoint variants to accommodate different memory and hardware requirements. AWS provides multiple deployment pathways, starting with Amazon SageMaker JumpStart for fully managed real-time endpoints. For users requiring Kubernetes-native control, Amazon SageMaker HyperPod Inference offers integration with SGLang or vLLM runtimes. These options allow developers to manage distributed GPU infrastructure, caching, and scaling according to their specific concurrency and latency targets. At launch, the model is accessible via SageMaker JumpStart in US East regions, enabling both researchers and developers to customize runtime environments.
aws