Reflection AI releases open-weight Beam coding model (501B total, 23B active parameters) under Apache 2.0, available on AWS SageMaker JumpStart

Reflection AI released Beam 1.0, a coding-focused mixture-of-experts model with 501B total parameters, of which 23B are active for any given token. Weights are on Hugging Face under the Apache 2.0 licence in three formats (BF16, FP8 and NVFP4). At launch it can be deployed through Amazon SageMaker JumpStart in two US regions (N. Virginia and Ohio) and through SageMaker HyperPod Inference using SGLang or vLLM. AWS published a sample deployment repository. The announcement comes from AWS's own blog, and it reports a 1M-token context window that Reflection has stated but that is not independently verified.

Teams that want a permissively licensed, large coding model they can host and modify themselves now have another option, with a ready deployment route on AWS. Apache 2.0 allows commercial use without the restrictions found in many "open" model licences, which matters for enterprises with data-control or customisation needs. The practical value is still unproven: the post gives no benchmarks, so nobody can yet say whether Beam beats existing open coding models or justifies the hosting cost of a 501B-parameter model. Even with only 23B parameters active, the full weights must be held in memory, so self-hosting needs serious GPU capacity.

The language about "open and accessible intelligence" and enabling "long-horizon tasks" is mission-statement packaging, not evidence of capability. The 1M-token context figure is a reported number with no shown test of how well the model uses that much text. The article is largely a deployment guide from a cloud vendor that benefits from hosting the model, so it says little about real coding quality.

1. 01Independent benchmark results on coding evaluations such as SWE-bench-style tasks, compared with existing open-weight coding models, within the next few weeks.
2. 02Evidence on the 1M-token context claim: third-party long-context tests showing whether quality holds up, plus real inference cost and latency on SageMaker.
3. 03Adoption signals: Hugging Face download counts, availability on other clouds or in more AWS regions, and any named enterprise users or fine-tuned derivatives.