Tech · NEWS EXPLAINER
Reflection previews Beam, with open weights planned later in October
The model targets coding and agentic work. Access and efficiency claims require different kinds of evidence.
Reflection introduced Beam, its first AI model, on October 5. The company is pitching it for businesses that want more control over their models.
Preview now, weights later
An early version is available to selected users, with a waitlist open. Reflection plans Apache 2.0 weights and supporting artifacts later this month.
Joining a waitlist is a request for access. It is not a completed download, and a promised license does not establish that the files are available.
Two counts, different meanings
Reflection lists 501 billion total parameters and 23 billion active per token.
Parameter counts describe the model, not a memory specification. Treating the smaller number as a hardware requirement would skip a necessary step: establishing what the entire deployment needs.
Estimate versus a bill
Its comparison estimates generation forward-pass compute from active parameters and generated tokens. It excludes prefill, context-dependent attention operations and serving overhead.
For a cost comparison, the useful question is what a complete workload costs to run. A narrower estimate cannot answer that without the missing operational information.
What still needs checking
Semafor reports that Reflection is positioning Beam as a workhorse for coding and agentic tasks. That is the developer's pitch, not an independent test result.
For an evaluation, access comes first; capability and operating expense then need their own evidence. A model announcement can describe an intended use without demonstrating the outcome.
