Overview
- Reflection publicly unveiled Beam on Monday as a 501-billion-parameter, sparse mixture-of-experts model that uses 23 billion active parameters, was pretrained on 23.8 trillion tokens, and supports a 1 million‑token context window.
- The company claims Beam matches Z.ai’s GLM-5.2 on advanced reasoning and coding benchmarks while using roughly three to four times less inference compute, though those results are self-reported and lack independent verification.
- Reflection plans to publish Beam’s full weights and a technical report later this month under a permissive Apache 2.0 license so developers and enterprises can download, modify, and self-host the model.
- Beam is central to Reflection’s ‘AI factory’ sales pitch that combines open weights, customer data, and dedicated Nvidia GB300 compute secured through multibillion-dollar deals with SpaceX’s Colossus and Nebius to let organizations build localized AI systems.
- The near-term tests for Beam’s impact are clear: third-party benchmarks, the exact licensing and distribution terms on release, and whether large enterprises sign up to self-host the model for cost savings and data sovereignty.