Pretraining could be completed in just 12 days, but Beam’s reinforcement learning phase consumed 10,500 GB300 chips over four full weeks. Misha Laskin, co-founder and CEO of Reflection AI @reflection_ai, broke down the training costs of the open model Beam. The company, founded in early 2026 by two former Google DeepMind researchers, raised a new round of funding in April with a $25 billion pre-money valuation. Beam has a total of 501B parameters and 23B activated parameters, having consumed 23.8 trillion tokens during pretraining; its weights were released under the Apache 2.0 license on October 16. The compute allocation between the two phases was inverted. Pretraining ran on 6,144 GB300 NVL72 chips—according to Laskin, this could now be completed in about 12 days with current infrastructure. The RL phase, however, used 10,500 identical chips continuously for four weeks, equating to roughly 40,000 GB300-weeks—four times the ~10,000 GB300-weeks used in pretraining—and consumed even more FLOPs.