Reflection positions Beam as a model for coding agents: the company’s Oct. 5 demos show it using tools to research documentation and build applications. On Oct. 16, Reflection made Beam’s weights downloadable on Hugging Face under Apache 2.0; the release offers BF16, FP8 and NVFP4 formats.
Weights move Beam from preview to self-hosting
The preview offered selected users an early-access API and set an expectation of weight release later in October. The Oct. 16 post now says developers can download and self-host the model, customize it, or use Reflection’s API and partner services. The weights license is Apache 2.0; that describes the model weights, not every service in the launch ecosystem.
Self-hosting requires Beam’s full weights; AWS says a full deployment needs multiple GPUs
Reflection says Beam has about 501 billion parameters in total and uses 23 billion to process each token. A self-hoster still has to store the full set of weights. AWS says a full Beam deployment requires multiple GPUs, with its BF16 configuration needing the most memory. The three formats provide different memory footprints, according to AWS. Beam is also a text model with a one-million-token context window, according to Reflection’s model card.
Reflection says its partners span managed endpoints, cloud, on-premises and sovereign infrastructure
Reflection says the release includes more than 40 partners and integrations, spanning managed endpoints, cloud services, on-premises systems and sovereign infrastructure. AWS lists Beam in SageMaker JumpStart in US East (N. Virginia) and US East (Ohio). Those are the two AWS regions named in its launch article; other partners provide separate access routes.
Image credit
Cover image: “Datacenter Server Racks (22370909788).jpg” by Carl Lender, licensed under CC BY 2.0. This generic data-center image does not depict Reflection or Beam hardware.
.jpg){kind=link}