Reflection announces Beam, a 501-billion-parameter open-weight model, but the weights are not out until later this month

Reflection, the US start-up that Axios reported on 4 October was about to release an open-weight model, published “Introducing Beam” on Monday 5 October. “We are introducing Beam, Reflection's first open-weight model,” the post says. It describes a sparse mixture-of-experts model with 501 billion total parameters, 23 billion of them active per token, built for coding, reasoning and agent work. Beam is text-only.

Nobody can download it yet. “This month, we will release the weights under an Apache 2.0 license, along with documentation and the full stack for running, evaluating, and fine-tuning the model,” Reflection writes, and it adds: “Beam is undergoing final red-teaming and evaluations.” Until then it is available to a select group of users through a waitlist.

Its pitch is efficiency against the leading Chinese open models. Reflection says that on advanced reasoning benchmarks Beam “achieves scores comparable to GLM-5.2 while using 3–4× less inference compute.” Its own table puts Beam slightly behind Z.ai's GLM-5.2 on Terminal Bench v2.1, at 80.1 against 81.0, and gives Beam 90.5 on GPQA Diamond.

TechCrunch reports that Beam was pretrained on 23.8 trillion tokens and has a 1-million-token context window, and notes that Reflection's performance claims have not been independently verified. It names Nvidia as a backer and compute supplier.

This answers the 4 October report that a release was coming: the model has been announced, but the open weights, the part that makes it open, are still a promise.

• Confirmed Beam is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active, announced 5 October. Reflection
• Claimed Weights to be released this month under Apache 2.0; Beam is still undergoing final red-teaming and evaluations. Reflection
• Claimed Scores comparable to GLM-5.2 with 3–4× less inference compute; Terminal Bench v2.1 80.1 against 81.0. Reflection
• Reported Pretrained on 23.8 trillion tokens with a 1-million-token context; claims not independently verified. TechCrunch