Reflection AI has released Beam, an open-weight model aimed specifically at agentic coding: tasks where a system has to maintain code over time, call external tools, read back the results of those calls, and carry out multi-step workflows rather than answer a single prompt. Amazon Web Services has published a companion guide showing how to deploy Beam using Amazon SageMaker, which is useful both as a practical how-to and as a window into what running a model of this size actually requires.

What is new

Beam is a sparse mixture-of-experts (MoE) model with 501 billion total parameters, of which 23 billion are active for any given token. Reflection reports an effective context length of 1 million tokens, meaning the model is designed to work with very large amounts of input text or code history in a single session. The weights are released on Hugging Face under the Apache 2.0 license, which is a permissive open-source license that allows commercial use, modification and redistribution. Beam 1.0 ships in three checkpoint formats: BF16, FP8 and NVFP4, each trading off memory footprint and hardware compatibility differently. At launch, Beam is also available directly through Amazon SageMaker JumpStart in two AWS regions, US East (N. Virginia) and US East (Ohio).

For readers unfamiliar with the terms: a mixture-of-experts model splits its parameters into many specialised sub-networks ("experts") and routes each token through only a handful of them, rather than through the entire network. This lets a model have a very large total parameter count, which can improve capability, while keeping the actual computation per token closer to that of a much smaller dense model. That is the gap between Beam's 501 billion total parameters and its 23 billion active parameters.

How it works

The AWS post is primarily a deployment guide, and it lays out two main ways to run Beam on SageMaker. The first is a SageMaker JumpStart real-time endpoint, described as the fastest path from the model catalog to a working HTTPS endpoint, with AWS managing the container, endpoint and invocation API. The second is SageMaker HyperPod Inference, which gives Kubernetes-native control over distributed GPU infrastructure, orchestrated through Amazon Elastic Kubernetes Service (EKS). Within the HyperPod path, developers can either use the AWS-managed JumpStart container configuration, or take direct control of the serving runtime using SGLang or vLLM, two popular open-source inference engines.

Because Beam is so large, hardware choice matters a great deal. According to the guide, the BF16 checkpoint has the largest memory footprint and needs a multi-node setup. For FP8, AWS recommends ml.p5.48xlarge instances with NVIDIA H100 GPUs, or ml.p5e.48xlarge and ml.p5en.48xlarge instances with H200 GPUs. For the more compressed NVFP4 format, the recommended instance is ml.p6-b200.48xlarge with NVIDIA B200 GPUs. The post stresses sizing the deployment not just for the model weights themselves but for checkpoint storage, the KV cache (the memory used to store intermediate attention computations during generation), runtime overhead, context length and expected concurrency, confirming estimates with real load tests rather than relying on back-of-envelope numbers alone.

On the HyperPod path, AWS highlights a few supporting features: an Inference Gateway that offers an OpenAI-compatible interface and routes requests across replicas, managed tiered KV caching, and intelligent routing. The post also describes model and image caching, which preloads weights onto node-local NVMe storage and pre-pulls the serving image onto cluster nodes so that scaling up does not require re-downloading large checkpoints each time.

Why it matters

Agentic coding systems - tools that read a codebase, make changes, run tests, interpret the output and decide what to do next - are one of the more demanding applications for large language models, because they require sustained reasoning across many steps rather than a single response. Reflection AI frames Beam as purpose-built for this kind of long-horizon workflow, and its reported 1 million token effective context length is relevant here: a model that can hold more of a codebase or a longer conversation history in context at once has more information available when deciding what to do next.

The open-weight release is also significant on its own terms. Because Beam's weights, rather than just an API, are published under Apache 2.0, developers and researchers can run the model on their own infrastructure, inspect it, fine-tune it, and integrate it into custom pipelines without depending on a vendor's hosted service. Reflection AI describes this as part of a broader mission to make intelligence open and accessible, spanning models, core software, production infrastructure and applications.

Limits and open questions

The source material is a deployment guide, not an independent evaluation, so it does not provide benchmark scores, accuracy figures, or comparisons against other coding models. Running Beam is also clearly not a casual undertaking: even the compressed NVFP4 format requires specialised B200 GPU instances, and the guide repeatedly stresses the need for GPU quota, capacity planning, and load testing before production use. The post does not say how Beam compares in cost or performance to closed, API-only coding models, nor does it provide details of the training data or methodology behind Beam.

How to try it

AWS provides two sample repositories for getting started. For the managed path, a SageMaker JumpStart sample lets a developer clone a repository, set up a Python environment, and run a single command that deploys an endpoint in a chosen checkpoint variant (BF16, FP8 or NVFP4), sends a test request, and tears the endpoint down afterward. For the Kubernetes-native path, a HyperPod Inference sample includes configuration templates for JumpStart, SGLang and vLLM deployments, along with a command-line tool to validate settings, deploy the model, and run a smoke test against a live endpoint. Both samples include matching cleanup instructions, which AWS recommends following promptly given the cost of the GPU infrastructure involved.

  • 501 billion total parameters, 23 billion active per token (sparse mixture-of-experts architecture)
  • Reported effective context length of 1 million tokens
  • Released on Hugging Face under the Apache 2.0 open-source license
  • Available in BF16, FP8 and NVFP4 checkpoint formats
  • Recommended hardware: H100 or H200 GPU instances for FP8, B200 GPU instances for NVFP4
  • Deployable via SageMaker JumpStart (managed) or SageMaker HyperPod Inference (Kubernetes-native, with SGLang or vLLM options)
  • Live in SageMaker JumpStart in US East (N. Virginia) and US East (Ohio) at launch

Sources