Reflection's Beam: a 501B open model on a fraction of the compute

Reflection AI, the Brooklyn startup founded by ex-DeepMind researchers, has unveiled Beam, its first frontier open-weight model. It is a mixture-of-experts design with 501B total parameters but only 23B active per token. Reflection trained it on 23.8 trillion tokens using a 6,144-GPU cluster.

The pitch is efficiency. Reflection says Beam gets close to Qwen 3.8-Max, a model with over 2 trillion parameters, on advanced reasoning and coding/agentic benchmarks. It claims to do so with roughly one-third to one-quarter of the hardware. These are the company's own numbers, and independent evals haven't landed yet.

There is a catch for anyone hoping to download it today. Weights are not out yet. Reflection says they will ship under Apache 2.0 later in October, along with a technical report, a model card and the full stack for running, evaluating and fine-tuning. The launch follows a funding round at a reported $25B valuation.

Why it matters: Chinese labs dominate open-weight models, and Beam is the best-funded Western attempt to close that gap on cost as well as quality.