Reflection AI introduces 501B parameter open model Beam

Reflection AI has introduced Beam, a new open-weight large language model, SiliconANGLE AI and MarkTechPost reported. As FTMQ SI reported earlier, the release occurred on October 5, 2026. The startup claims the model matches leading Chinese open models on reasoning benchmarks at lower compute costs, TechCrunch AI reported. [2][3][4][13]

The model features a 501-billion parameter sparse Mixture-of-Experts architecture with 23 billion active parameters, according to MarkTechPost. MarkTechPost reported that Reflection AI says Beam matches GLM-5.2 on reasoning tasks while using three to four times less inference compute. SiliconANGLE AI noted that Reflection AI trained the model using Nvidia GB300 NVL72 appliances containing 72 graphics cards each, rented through a reported 6.3 billion dollar deal with SpaceX. [2][3]

Reflection AI was founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou to focus on automated software development tools, Wikipedia describes. SiliconANGLE AI added that the launch comes several months after the startup raised funding at a 25 billion dollar valuation. TechCrunch AI reported that the Brooklyn-based company is targeting enterprises and sovereign nations with a product called AI factories, which allows institutions to train customized local systems on proprietary data. [2][4][10]

Reflection AI plans to release the Apache 2.0 open-weight code later in October 2026, MarkTechPost reported. [3]