Reflection AI has launched Beam, an open-source large language model featuring 501 billion parameters. The company previously secured a valuation of $25 billion and entered a $6.3 billion agreement with SpaceX to utilize Nvidia GB300 NVL72 hardware for training purposes.
The model was benchmarked against GLM-5.2 and compared to Qwen 3.8-Max. According to the developer, Beam achieves comparable performance while utilizing significantly less hardware. It is the first open-source model from a U.S. startup to reach such performance levels.
The training process for Beam Base involved 6,144 graphics cards and 23.8 trillion tokens, with development completed in under four weeks. The company utilized custom filters to process code-heavy datasets and implemented midtraining to improve reasoning and context window capabilities.
The final training phase utilized 10,000 GB300 graphics cards to run 1.3 billion reinforcement learning sandboxes. This stage also lasted four weeks, with a median error recovery time of eight minutes across 71 recorded malfunctions.
Beam is currently accessible through an early access program. Reflection AI intends to release the model weights, documentation, and fine-tuning tools to the public later this month.
