Reflection AI launches open-weight model Beam at lower compute cost
Reflection AI has officially launched Beam, its first frontier open-weight AI model, claiming it matches leading Chinese models like GLM-5.2 on reasoning benchmarks while using 3-4x less inference compute. The 501B-parameter MoE model (23B active) was trained on 23.8T tokens including an OCR pipeline over hundreds of millions of PDFs, and features a 1M token context window. It targets enterprises, public sector, and sovereign nations, with plans for 'AI factories' allowing customization on proprietary data. Reflection has raised ~$4.7B from backers including Nvidia and Sequoia, and signed compute deals worth over $7B with SpaceX and Nebius for GB300 chips. The company reportedly spends $150M/month on Colossus compute. Independent analyses place Beam around GLM-5.2 level, below DeepSeek V4 Flash on some benchmarks. Beam's weights (under Apache 2.0) and technical details will be released this month via hyperscalers and neoclouds.
- •Reflection AI launched Beam, a 501B-parameter open-weight MoE model with 23B active parameters, trained on 23.8 trillion tokens and a 1 million token context window.
- •Reflection claims Beam matches Z.ai's GLM-5.2 on reasoning benchmarks while using 3-4x less inference compute, and scores 80.9 on SWE-bench Verified.
- •Reflection has raised approximately $4.7 billion from backers including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, and signed compute deals worth over $7 billion with SpaceX and Nebius for Nvidia GB300 chips.