Reflection AI unveils Beam to challenge Z.ai's GLM-5.2, 501B model
Reflection AI, founded in 2024 by two former Google DeepMind researchers, officially unveiled Beam, its first frontier, open-weight AI model.
Built with 501 billion parameters and pre-trained on 23.8 trillion tokens, Beam is designed to take on big Chinese models like Z.ai's GLM-5.2, but promises similar power with much lower computing costs.
Beam reportedly uses 3-4x less compute
Beam focuses on text tasks like reasoning and coding, using a "mixture-of-experts" approach that keeps things efficient.
It reportedly uses 3-4x less inference compute than rivals. The team pre-trained it on 23.8 trillion tokens.
Backed by $4.7 billion from names like NVIDIA and Sequoia Capital, Reflection AI has also inked multi-billion-dollar deals for top-tier chips through 2029, aiming to help companies build custom AIs.
Expect Beam's details and weights to drop later this month for anyone curious or ready to experiment.