Reflection AI introduced Beam, a 501-billion-parameter open-source LLM, a few months after raising funding at a 6.3 billion deal with SpaceX to rent Nvidia GB300 NVL72 appliances used to train Beam.
Reflection benchmarked Beam against GLM-5.2 (~250B more parameters) and found Beam performs some tasks better using one-third to one-fourth the hardware. Beam approaches Qwen 3.8-Max performance despite Qwen having over 2 trillion parameters.
Beam is the first open-source model from a U.S. startup demonstrating comparable performance to leading Chinese open models. Free LLMs still trail frontier closed models like Anthropic’s Claude Fable 5.1.
Beam available through early access; weights and fine-tuning tools promised later in October.