Reflection AI releases open weights for Beam, a 500B parameter coding model

Reflection AI released the weights of Beam, its first large language model, on Hugging Face under the permissive Apache 2.0 license. The company also published technical resources and post-training recipes so developers can customize the model for their own needs.

Beam has about 500 billion parameters and is built for agentic coding and long tasks: maintaining code, using tools, reading results and completing multi-step workflows. Reflection pitches it as token-efficient and aimed at teams that want to run a powerful model on their own hardware, infrastructure or cloud. Support landed in a new Hugging Face Transformers release, and AWS published a guide to deploying it through SageMaker.

It is one of the largest open-weight coding models from a US company, arriving in a week when Chinese labs kept shipping cheaper models of their own.

Open weights also mean companies can fine-tune Beam on their own codebases and keep that code off third-party servers, which matters for regulated industries and anyone with strict data rules.

Our take

A US-made, Apache-licensed frontier-scale coding model is the hedge many enterprises wanted. The catch is the hardware bill to run 500 billion parameters yourself.