Reflection Beam

Beam (announced Oct 5, 2026) is the first model from Reflection AI, the Nvidia-backed startup founded by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou. It is a text-only sparse mixture-of-experts model: 501B total parameters, 23B active per token, pretrained on 23.8T tokens, then RL-tuned in what Reflection calls one of the largest open-lab RL runs (100M+ rollouts on 10.5K NVIDIA GB300 GPUs, graded in ~1.3B sandboxes). Its pitch is intelligence-per-token: matching Z.ai's GLM-5.2 on reasoning at 3–4× less inference compute. Important: the weights are NOT out yet. Beam is in early access via a waitlist while it finishes red-teaming; Reflection says weights, technical report, model card, and tooling release later in October 2026 under Apache 2.0. All benchmark numbers to date are Reflection's own and have not been independently verified.

Provider-reported only — none independently verified (TechCrunch explicitly noted this): Terminal-Bench v2.1 80.1; SWE Bench Pro v2-Hard 77.2 (Thinking Machines Inkling: 56.9); SWE Bench Pro v1 65.5; SWEBench Verified 80.9; GPQA Diamond 90.5; DeepSWE v1.1 44.4; Humanity's Last Exam 36.2 (no tools). Reflection's own table shows newer Chinese open models (Z.ai GLM-5.3, Moonshot Kimi K3) ahead on most tests.

Independent: no third-party evals exist yet. Independent news coverage is extensive (Axios scoop pre-announcement, FT, TechCrunch, Business Insider, DataCamp guide, explainx.ai deep dive flagging what is unverified).