Reflection unveils Beam AI preview, with open weights promised this month

Beam begins with limited early access

Reflection AI introduced Beam on October 5 as its first model intended for an open-weight release. The company describes it as built for coding, reasoning and tasks in which AI systems use tools to carry out multi-step work. Reuters independently reported the announcement.

There is an important distinction between the preview and full release. Reflection says an early version is available to selected users through a waitlist. It plans to release the model weights, documentation and tools for running, evaluating and fine-tuning Beam later this month under an Apache 2.0 licence. The announcement does not establish that those promised materials are already publicly downloadable.

A large model using a smaller active network

The company says Beam has 501 billion total parameters but uses 23 billion active parameters, a design known as a sparse mixture of experts. Parameters are the model’s learned numerical settings. Activating only part of the network is intended to reduce the computing work required compared with using the entire model for each step.

Reflection reports competitive results against other models on coding and agent tasks. Those are vendor evaluations, not independent INCE NEWS testing. They do not establish how Beam will perform on every real deployment or how much customers will ultimately pay to run it.

The release will make scrutiny possible

Reuters places the launch within growing competition between US AI companies and Chinese open-weight models. Reflection also says it will publish a technical report and safety evaluations. Wider access to the promised materials will be important for testing the company’s claims; an announcement and benchmark table are only the starting point for that assessment.