AWS said in a 16 October post that Reflection AI has introduced Beam, an open-weight model for coding and multi-step tasks. Open-weight means the trained files are public, so teams can run them on their own. Reflection releases them on Hugging Face under an Apache 2.0 license, in BF16, FP8 and NVFP4 versions. At launch, AWS says, Beam is on SageMaker JumpStart in N. Virginia and Ohio.
Beam is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active ones. AWS says the total shapes checkpoint storage, weight memory, sharding and parallelism. The active count influences compute per token. AWS says the total requires several GPUs. BF16 needs the most weight memory, so it calls for a multi-node setup. For FP8, AWS names H100 or H200 servers. For NVFP4, it names B200 servers. Reflection reports an effective context length of 1 million tokens. AWS gives no prices or cost comparisons.
Memory and GPU layout follow the 501 billion total, and the file format changes the hardware needed. That is what AWS's post supports. Our reading is that buyers should price the cost of reserving that capacity, since AWS gives no figures. AWS also advises checking GPU quota and capacity for the chosen server type.