Beam’s model weights are now available, beyond its earlier preview

Reflection’s Beam now has downloadable model weights under Apache 2.0, a concrete change from its earlier announcement promising them later. The text model has 501 billion total parameters, with 23 billion active per token, and a documented one-million-token context window. Reflection provides BF16, FP8 and NVFP4 versions. The BF16 repository lists roughly one terabyte of files, so the active-parameter count should not be mistaken for its storage footprint.

Reflection

Why it matters

Available weights make inspection and deployment possible in ways that an early-access announcement alone does not.

More context

The official collection links all three weight formats, and developer documentation covers hosted API access and running your own deployment. Inspect those requirements before downloading; try a hosted evaluation first if hardware is limited. The exact original date of public weight availability was not established.

Keep in mind

The technical report is still marked forthcoming. Context capacity and benchmark claims come from Reflection’s documentation; neither deployment costs nor advertised performance were independently measured for this card.

• Weight files and sizes
• Official model collection
• Deployment and API documentation
• Original October 5 preview

Original read