Reflection introduces Beam, with open weights promised later this month

Reflection has introduced Beam, a new AI model focused on coding, reasoning and software-agent tasks. The company is describing it as an open-weight model, although the weights are not being released immediately.

Beam uses a mixture-of-experts design with 501 billion total parameters and 23 billion active parameters. In simple terms, the full model is large, but only part of it is activated for each token it processes.

Reflection says final safety testing and evaluations are still underway. Early-access registration is available, with the weights, technical report and supporting developer materials scheduled for release later this month.

The release timing is the part I would pay attention to. An announcement and a model that developers can download and test are not the same milestone. The company’s performance comparisons also remain its own claims until independent testing provides a clearer picture.

For developers, Beam offers another model to evaluate rather than an automatic replacement for an existing tool. The useful questions will be how reliably it handles real tasks, what it costs to run and what its eventual licence permits.