Overview
- Reflection publicly unveiled Beam on Monday as a sparse mixture-of-experts model with 501 billion total parameters that activates about 23 billion parameters per token to reduce run-time cost.
- The company reports Beam was pretrained on roughly 23.8 trillion tokens, uses large-scale reinforcement learning and supports a 1 million token context window to improve multi-step reasoning and agentic tasks.
- Reflection claims Beam matches or approaches leading Chinese open models on coding and reasoning benchmarks while using roughly three to four times less inference compute, but those results are self-reported and not yet independently verified.
- The startup is positioning Beam inside an "AI factory" product for enterprises and governments, has Nvidia backing and multi-billion dollar GB300 compute deals with SpaceX (Colossus) and Nebius, and intends to publish the model weights under an Apache 2.0 license later in October.
- The release sharpens a Western open-weight push that expands options for firms seeking local control and lower cost while raising safety, misuse and oversight questions that depend on the results of Reflection’s final red‑teaming and outside benchmarks.