Reflection previews Beam; downloadable weights are promised for October
Reflection has opened a selective preview of Beam and promises Apache 2.0 weights this month; its efficiency claim remains an estimate.
Reflection’s first model, Beam, is entering a selective preview, with public weights and technical documentation promised later in October.
Its announcement gives no exact release day. The prospective benefit is another model that organisations could operate themselves once the promised release arrives. [1]
The company reports 501 billion parameters in total, with 23 billion active for a token. Its claimed three-to-fourfold compute advantage uses an approximation that leaves out prompt processing, context-dependent attention and serving overhead. That does not establish a comparable reduction in bills or waiting time. [1]
In an interview with Semafor, CEO Misha Laskin described customers seeking control over their own AI systems. That is a deployment argument, separate from winning every benchmark. The interview independently establishes the company’s positioning; it does not reproduce its tests. [2]
Sources & details 2 sources
Discuss in your voice app
Take the reporting, visual explanations and original sources into a conversation. Copy the brief, paste it into your preferred assistant, then start its voice mode.
Paste the brief into your voice app to discuss this reporting.Revised 9 October 2026. Copy edited 9 October to remove commentary about graphic design; reported claims, qualifications and source dates preserved.
Reporting sources
- Reflection · October 5 announcement
Original announcement, architecture, compute-estimation omissions and selected preview; weights and Apache 2.0 licence remain promised. Reopened after the collection barrier.
- Semafor · Interview with Misha Laskin
October 5 original reporting and interview; corroborates the announcement and customer positioning, not benchmark performance.
Images & figures
The active portion is 23/501 of the total parameter count, about 4.6%. A sparse model uses only part of its parameter pool for a token; these counts do not establish its total memory needs or end-to-end cost. Company-reported architecture.
Euangelion · original explanatory graphic · Source · Usage terms