Reflection AI's Beam, a 501B open-weight coding model, arrives on AWS

What Beam is

According to AWS, Beam is a sparse mixture-of-experts model for resolving software issues, modifying code, calling tools, and acting on intermediate results. It is aimed at long-horizon work such as maintaining code, using tools, interpreting results, and completing multi-step workflows.

The model has 501 billion total parameters and 23 billion active parameters. The total count shapes checkpoint storage, weight memory, sharding, and parallelism, while the active count influences per-token compute. Reflection reports an effective context length of 1 million tokens.

Reflection releases Beam's weights through Hugging Face under the Apache 2.0 license. Beam 1.0 comes in BF16, FP8, and NVFP4 checkpoint variants, which differ in memory footprint and hardware compatibility. The open-weight release gives developers and researchers control over the runtime, deployment environment, and customization process. Reflection says its mission is to make intelligence open and accessible through open model weights, published research, and open source software.

Two ways to deploy on AWS

AWS describes a managed path and a Kubernetes-based path.

• SageMaker JumpStart real-time endpoint: a fully managed model package, endpoint, and invocation API, with no Kubernetes to operate. AWS recommends starting here unless a workload needs a specific inference runtime or configuration.
• SageMaker HyperPod Inference: for Kubernetes-native control over distributed GPU infrastructure, on a cluster orchestrated by Amazon EKS. It can run the AWS managed JumpStart container, or SGLang or vLLM when teams need to control the serving image, runtime arguments, quantization, or distributed execution.

For the JumpStart path, AWS points to a sample in its sagemaker-genai-hosting-examples repository. A single variant-aware command deploys an endpoint, sends a short request, prints the response, and cleans up. NVFP4 is the default if no variant is specified.

On HyperPod, the Inference Gateway provides an OpenAI-compatible interface and routes requests across replicas. HyperPod Inference also supports managed tiered KV caching and intelligent routing. For production and model caching, AWS advises copying the checkpoint to Amazon S3 or Amazon FSx, while a pinned Hugging Face revision suits development and experimentation.

Hardware and capacity planning

Because of its size, Beam requires a multi-GPU deployment. BF16 has the largest weight-memory requirement, so AWS advises a multi-node configuration with enough aggregate GPU memory for weights and runtime overhead.

• FP8: ml.p5.48xlarge with NVIDIA H100 GPUs, or ml.p5e.48xlarge and ml.p5en.48xlarge with H200 GPUs.
• NVFP4: ml.p6-b200.48xlarge with NVIDIA B200 GPUs.

AWS recommends calculating memory for weights, KV cache, and runtime overhead for one replica, then confirming the estimate with a load test. Teams should check service quotas and capacity in their chosen Region, and can use on-demand capacity or a SageMaker training plan as a short-term reservation for scheduled work. Time should be allowed for image transfer, weight download, shard loading, and runtime initialization.

What it means for developers

For teams building software engineering and agentic systems, the release pairs openly licensed weights with a documented route to run them on AWS infrastructure. Those who want minimal operations can use the managed JumpStart endpoint, while those who need to tune serving images, quantization, caching, and routing can use HyperPod. Availability at launch is limited to two US Regions for JumpStart, and the hardware requirements show that running a model of this scale demands substantial GPU capacity.