Reflection AI debuts open-source Beam model with 501B parameters
SiliconANGLE Maria Deutscher ● Covered by 3 sources
Reflection AI shipped Beam, a 501B-parameter open model. It says it rivals much bigger Chinese models while using far less hardware.
Based on reporting by SiliconANGLE, Maria Deutscher — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Reflection AI has released Beam, an open-source large language model with 501 billion parameters. The startup says it trained the model on Nvidia GB300 NVL72 systems rented from SpaceX, and that each of those appliances carries 72 graphics cards. That is a lot of silicon even by AI standards, but Beam’s pitch is efficiency: the company says it can match or beat some tasks from much larger models while using only about a third to a quarter of the hardware.
The comparison set matters here. Reflection AI benchmarked Beam against GLM-5.2, another open-source model, and says Beam comes close to Qwen 3.8-Max, which has more than 2 trillion parameters. That puts a U.S. startup in a spot that has mostly belonged to Chinese companies in the open-source LLM race. The company is framing Beam as the first open-source model from a U.S. startup to show comparable or better performance in that company.
The training process was built in layers. Reflection AI says it started with a relatively small prototype, then scaled through a series of larger models until it reached Beam Base, the foundation model underneath Beam. Beam Base was trained on 23.8 trillion tokens from the public web and commercial sources, with a heavy dose of software code. To keep the code data clean, the company built custom filters for each programming language to cut out low-quality files.
Speed is part of the story too. Beam Base was developed in under four weeks. After that came midtraining, which extended the context window and improved reasoning, and then a reinforcement learning phase that ran on 10,000 GB300 graphics cards and 1.3 billion sandboxes. Those sandboxes were tuned for code generation, web search and AI agents. Reflection AI says the final reinforcement learning stage took four weeks, and that its cluster recovered from 71 errors with a median delay of eight minutes. Beam is now in early access, with weights, documentation and fine-tuning tools due later this month.
My take — AI-written commentary, not fact-checked reporting
This is the kind of open model story that actually matters: not a louder benchmark victory lap, but an efficiency play with enough scale to annoy the incumbents. The open-source field has spent too long pretending bigger alone is a strategy; Beam says the real flex is getting more out of less. The U.S. finally has a model that can stop gawping at Chinese releases for five minutes.
Read more about this at: SiliconANGLE
