Reflection AI releases open weights and post training recipes for Beam model

Startup Reflection AI Inc. has made the weights of its first large language model, Beam, available on Hugging Face, SiliconANGLE reported. SiliconANGLE noted that the startup also released technical resources that allow software developers to customize the algorithm. Unite.AI and shattered.io both reported that Reflection AI released the open weights under an Apache 2.0 license. [1][3][5][6]

Originally introduced earlier in October, Beam features approximately 500 billion parameters, SiliconANGLE reported. SiliconANGLE added that the model approaches the performance of algorithms four times its size on certain benchmarks while using significantly less computing power and memory. VentureBeat reported that Reflection AI aims to undercut Chinese open models by using fewer tokens. [1][3][4]

As FTMQ SI reported earlier, Beam is a 501-billion parameter sparse Mixture-of-Experts model with 23 billion active parameters. It targets coding and agentic tasks while matching GLM-5.2 reasoning performance using three to four times less inference compute. Reflection AI trained the model on rented Nvidia GB300 NVL72 appliances. [9][10]

Reflection AI is an American artificial intelligence company founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou. The company focuses on tools that automate software development and foundation models. Misha Laskin serves as the chief executive officer of the company. [7][8]