What changed
On October 16, Reflection announced that Beam's weights were available, moving beyond the earlier preview. The model targets coding and multi-step tool-using tasks. Its model card describes a text-in, text-out mixture-of-experts model with 501 billion total parameters, 23 billion active per token, and a reported one-million-token context window. BF16, FP8, and NVFP4 checkpoint variants are released under Apache 2.0. These are provider descriptions, not our measurements.
The AWS route has hardware requirements
AWS's October 16 deployment guide lists Beam in SageMaker JumpStart in US East (N. Virginia) and US East (Ohio) at launch. Its documented configurations use multiple GPUs. Check which instance types and checkpoint formats the deployment guide supports. Longer contexts need more memory; check the context length supported by the deployment you choose. GPU quotas, available capacity, checkpoint storage, and runtime overhead also matter. An open model license does not make cloud hosting free. We have not downloaded or run Beam.
Our take
Editorial analysis: treat the released weights and documented hosting route as useful access changes, then check the exact configuration before committing resources. Only about 23 billion parameters are active for each token; storage and weight memory depend on the full model and checkpoint format. For a small experiment, compare supported hardware, the context length you actually need, and the cost of a limited test. If you use Beam as a coding agent, keep permissions narrow and inspect proposed code changes before applying them.