Reflection Beam open model lands on AWS with SageMaker support

Reflection AI introduces Beam, an open-weight sparse mixture-of-experts model for agentic coding and long-horizon tasks. It has 501B total parameters with 23B active, a reported 1M-token effective context, and weights released on Hugging Face under Apache 2.0 in BF16, FP8, and NVFP4 checkpoint variants. Beam is now available on Amazon SageMaker JumpStart in US East (N. Virginia) and Ohio. Two deployment paths exist. Teams on SageMaker can use JumpStart real-time endpoints with AWS-managed containers, models, and endpoints for minimal ops. Those needing Kubernetes-native control can use SageMaker HyperPod inference with JumpStart containers or their own SGLang or vLLM runtimes, controlling serving images, runtime parameters, quantization, and distributed execution. For memory, FP8 suggests ml.p5.48xlarge (H100) or ml.p5e/ml.p5en.48xlarge (H200), while NVFP4 maps to ml.p6-b200.48xlarge (B200). For developers running open coding models on AWS, Beam pairs weights with clear GPU instance mappings and HyperPod support for model caching, tiered KV cache, and inference gateway routing, reducing multi-node deployment trial and error.