Reflection AI has unveiled Beam, its first large open-weight AI model. With this model, the U.S. company aims primarily to compete with powerful open models from China. Reflection claims that Beam delivers comparable performance while requiring significantly less computing power.
Beam is a mixture-of-experts model with 501 billion parameters, with 23 billion active during use. According to Reflection, the model was trained on 23.8 trillion tokens. The architecture is designed to run Beam relatively efficiently despite its size.
According to Reflection, Beam performs on par with major Chinese open models on various benchmarks. The company cites GLM-5.2, among others, as a point of comparison. TechCrunch does note, however, that the presented benchmark results have not yet been independently verified.
Less computing power
Reflection places a strong emphasis on efficiency with Beam. The company states that the model can be trained and used with less computing power than many competing frontier models. In doing so, it aims to address an increasingly important issue: the rapidly rising costs of training and running large AI models.
This positioning aligns with Reflection’s ambition to offer an American alternative to open models from China. Chinese developers, in particular, have released many powerful models in recent years whose weights are publicly available.
Reflection now has ample capital and computing power to take on that competition. According to TechCrunch, the company has raised approximately $4.7 billion and was valued at $25 billion pre-money in its most recent funding round. It also signed agreements worth more than $7 billion for computing power with SpaceX and Nebius, which will utilize Nvidia GB300 chips, among others.
Model weights to follow
Although Reflection has now officially announced Beam, the model weights are not yet available. The company says it will publish them later this month, along with more detailed technical information. Only then will it become easier for external researchers and developers to independently verify the performance and efficiency claims.
With Beam, Reflection is thus entering an increasingly crowded field of open-weight models. The key question is not only how Beam performs on benchmarks, but also whether the promised efficiency holds up in practice.