Reflection AI launches 501B-parameter open-source Beam model

Reflection AI has introduced Beam, an open-source large language model with 501 billion parameters, trained using Nvidia GB300 NVL72 systems rented from SpaceX in a $6.3 billion deal. Benchmarked against GLM-5.2, Beam outperforms it on some tasks using one third to one fourth the hardware, and approaches the performance of Qwen 3.8-Max, a model with over 2 trillion parameters. Trained on 23.8 trillion tokens in under four weeks, Beam is notable as the first U.S. open-source model to match leading Chinese-developed LLMs. Weights and fine-tuning tools will be released later this month.