Reflection AI Beam-1.0 launches with day-one serving across clouds

Reflection AI’s Beam launch pairs a large open-weight agentic coding model with ready-made serving paths instead of leaving deployment to adopters. Beam is a 501B-parameter mixture-of-experts model with 23B active parameters per token, and supports reasoning, tool calling, structured outputs, and a 1M-token context window. vLLM, AWS, Google Cloud, and Vercel each provide a route to run or access it, spanning self-managed GPUs, Kubernetes-native infrastructure, managed endpoints, and an API gateway.

• Beam-1.0 has 501B total parameters and 23B active per token; vLLM supports tensor, data, and expert parallelism plus reasoning, tool calling, and structured outputs.
• vLLM reports 2.71× lower batch-1 decode latency and 1.87× faster prefill than its first implementation; DFlash2 speculative decoding reached 708 tokens/s on TP8 with an acceptance length of 3.7.
• AWS offers Beam through a SageMaker JumpStart real-time endpoint or SageMaker HyperPod Inference, with JumpStart, SGLang, or vLLM for Kubernetes-native control of distributed GPU infrastructure.
• Google Cloud describes up to 5× higher throughput and 8× lower latency on AI Hypercomputer using llm-d and NVIDIA Dynamo.
• Vercel AI Gateway exposes Beam with one API key, spend tracking, request traces, adjustable reasoning effort, and limited-time free access; the weights are released under Apache 2.0.

Bottom line If you build or operate coding agents, benchmark Beam on your intended serving path now; otherwise, skim unless this stack is relevant.