A model developer with public artifacts
Reflection AI develops language models and software for using them. Its Beam release provides a concrete example: a mixture-of-experts model with 501 billion total parameters and about 23 billion active per token, accompanied by an API and coding interface. Those artifacts establish a role beyond its earlier fundraising announcements. [2][1]
The Beam announcement on October 5, 2026 initially offered early access and said weights would follow. That date should not be described as the public weight release. [2]
Weight availability followed the preview
On October 16 the company announced downloadable Beam weights under Apache 2.0, a public API beta and the Mirror command-line coding interface. Hosted API access and operating the weights directly are separate deployment choices. The later release supplies the evidence that the promised public checkpoint became available. [1]
Reflection's capability comparisons and training descriptions are its own account. The cited releases establish model identity, licensing and access stages; they do not independently demonstrate superiority over competing models or correctness on a particular engineering task. [2][1]
Sources
- Beam is now available
Reflection AI · Oct 16, 2026
- Introducing Beam: Reflection’s 501B open-weight model
Reflection AI · Oct 5, 2026