Overview
- Reflection publicly unveiled Beam on Monday, Oct. 5, describing it as a 501-billion-parameter sparse mixture-of-experts model that activates about 23 billion parameters per token and was pretrained on roughly 23.8 trillion tokens.
- The company says a four-week reinforcement-learning run used about 10,500 NVIDIA GB300 GPUs, produced more than 100 million rollouts and used roughly 1.3 billion sandboxes to train Beam’s agentic and coding skills.
- Reflection claims Beam matches or approaches top Chinese open models on advanced reasoning and coding benchmarks while using roughly 3–4 times less inference compute, but those performance and efficiency figures are self-reported and not yet independently verified.
- Reflection has locked in large multi-billion-dollar compute commitments with SpaceX (Colossus) and Nebius and has Nvidia backing, which it says enables its commercial pitch for customer‑hosted “AI factories” and early sovereign tests such as a Shinsegae partnership in South Korea.
- The company plans to publish Beam’s full weights and a technical report under an Apache 2.0 license later in October, a step that will allow broad commercial reuse but also raises oversight, safety testing, and regulatory questions as final red‑teaming and independent benchmarks remain pending.