Reflection AI has introduced Beam, an open-weight large language model aimed at agentic coding and long-horizon tasks such as maintaining code, calling tools, interpreting results, and completing multi-step workflows. Beam is a sparse mixture‑of‑experts architecture with 501 billion total parameters and 23 billion active parameters, and it supports an effective context length of one million tokens. The model is released under the Apache 2.0 license on Hugging Face and is available in BF16, FP8, and NVFP4 checkpoint variants, each with different memory footprints and hardware compatibility. Deployment options on AWS include Amazon SageMaker JumpStart for a fully managed real‑time endpoint, and SageMaker HyperPod Inference for Kubernetes‑native control over distributed GPU infrastructure. Users can choose to run Beam with the managed JumpStart container, or customize the runtime with SGLang or vLLM. Multi‑GPU deployment is required for the full model; recommended instance types include ml.p5.48xlarge with NVIDIA H100 GPUs for FP8, ml.p5e.48xlarge and ml.p5en.48xlarge with H200 GPUs, and ml.p6-b200.48xlarge with NVIDIA B200 GPUs for NVFP4. The HyperPod path exposes cluster and serving runtime controls, allowing users to manage caching, routing, and scaling, while the JumpStart path packages the model, container, and endpoint behind a managed workflow. Beam’s open‑weight release enables developers and researchers to tailor the model to specific workloads and infrastructure, fostering broader experimentation and adoption of large language models in software engineering and agentic systems.
Introducing Reflection Beam: an open-weight model for agentic coding on AWS
Reflection AI has released Beam, an open-weight large language model designed for agentic coding and long-horizon tasks such as code maintenance, tool use, and multi-step workflows. Beam is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active parameters, offering an effective context length of one million tokens. The model is available in BF16, FP8, and NVFP4 checkpoint variants and can be deployed on AWS via Amazon SageMaker JumpStart or SageMaker HyperPod Inference, with options to use SGLang or vLLM for custom runtimes. Beam’s weights are released under the Apache 2.0 license on Hugging Face, and the model can be run on a range of GPU instance types such as ml.p5.48xlarge, ml.p5e.48xlarge, ml.p5en.48xlarge, and ml.p6-b200.48xlarge.
Why it matters
The open‑weight release gives developers control over runtime and deployment, supporting more flexible and scalable AI solutions for software engineering tasks. It also expands the ecosystem of large models available on AWS, encouraging experimentation and innovation in agentic AI.