Reflection AI introduces Beam but has not released its weights
Reflection named and detailed its first open-weight model, but access is limited while final safety tests continue.
A named model after the reported plan
Reflection AI introduced Beam on October 5, supplying details that were absent from the previous day’s Axios report about its open-weight plans. Reflection calls Beam a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active parameters. It says the system targets coding, reasoning and agentic work. Fortune reported the announcement at 3 p.m. Eastern time and noted that Reflection’s comparisons with competing models come from the company’s own tests. Those results are claims from Reflection, not independent evaluations of publicly downloadable weights.
Access is still limited
Reflection says Beam is undergoing final red-teaming and evaluation. Interested users can register for early access, and the company says a select group can use a preview. It plans to release the weights under an Apache 2.0 license later in October, alongside a technical report, model card and developer materials. That schedule is a stated plan, not a completed open-weight release. Until those artifacts are published, outside researchers cannot reproduce the benchmarks or fully assess the safety and operating characteristics described in the announcement.