Reflection AI model releases

Beam is Reflection AI's first public model, debuted on 2026-10-05 as a frontier-class open-weight text LLM positioned to rival leading Chinese open models at lower inference cost. It is a sparse Mixture-of-Experts model with 501B total / 23B active parameters (52 layers), trained on ~23.8T tokens, with a 1,048,576-token (1M) pretrained context (API beta requests capped at 262,144 tokens). Text in / text out (not multimodal). Weights are promised under Apache-2.0 later in October 2026, with distribution via hyperscalers/neoclouds and open-source library integrations; until then access is through an OpenAI-compatible API beta (model id Beam-501B-A23B at api.reflection.ai). Reflection reports SWE-bench Verified 80.9, Terminal-Bench v2.1 80.1, SWE-Bench Pro v2-Hard 77.2, DeepSWE v1.1 44.4 and Humanity's Last Exam 36.2, and claims it matches Z.ai's GLM-5.2 on reasoning while using 3-4x less inference compute (it trails Kimi K3 and DeepSeek V4.1 Flash on some agentic-coding tasks). All figures vendor-reported and unverified at launch.