The weights are out
Reflection released Beam’s weights on Hugging Face under an Apache 2.0 licence on 16 October, eleven days after announcing the model without them. The release also includes the tokenizer, a reference implementation, a model card with final benchmarks and a separate safety technical report. Weights come in BF16, FP8 and NVFP4, alongside a speculative decoding drafter model.
Beam is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active. Apache 2.0 permits commercial use, modification and redistribution subject to its notice requirements, which puts it at the permissive end of the open-weight spectrum rather than under the bespoke licences several rival releases carry.
What it takes to run
The three precisions exist because they land on different hardware. Amazon’s own post, published the same day, says Beam is available through SageMaker JumpStart in US East (N. Virginia) and US East (Ohio), and recommends ml.p5.48xlarge H100 instances or ml.p5e and ml.p5en H200 instances for FP8, and ml.p6-b200.48xlarge B200 instances for NVFP4. BF16 needs a multi-node setup because of its memory footprint. AWS repeats Reflection’s figure of an effective context length of one million tokens, and attributes it to Reflection.
Away from the cloud, Reflection lists self-hosting and partner endpoints, with on-premises and air-gapped deployment through Dell and HPE. That is the configuration enterprise and government buyers ask for, and the thing an API-only model cannot offer at any price.
The training bill
The most striking numbers in the post are not about the model but about what it cost to make. Reflection says it ran roughly 1.3 billion sandboxes across reinforcement learning training, evaluation and ablations. The final reinforcement learning run peaked at 883,006 virtual CPUs spread over 23 clusters on Google Cloud and AWS.
Benchmarks, and what is missing
The availability post carries no benchmark table of its own, pointing readers to the model card. Every figure there is Reflection’s, produced by Reflection. The company’s pitch since the 5 October announcement has been inference efficiency rather than peak capability — it said then that Beam matched GLM 5.2 while using a quarter of the inference compute.
What the availability post offers instead is anecdote. A pagoda built by one user took 12,600 tokens and 117 seconds. A Snake implementation finished in 2 minutes 19.7 seconds against 4 minutes 56.4 seconds for a comparison run. Those are demonstrations, not measurements, and Reflection presents them as such.
Pricing is not stated anywhere in the post. The platform is described as free to start during the public beta, which leaves the cost of using it through Reflection rather than running it yourself an open question.
What to watch
Reflection says self-distillation recipes, which it describes as raising token efficiency on a user’s own tasks, are coming “in the coming days”. The more consequential test is whether independent evaluators re-run the model card’s numbers now that anyone can download the weights and try — which is, after all, the point of releasing them.
