On 5 October, Reflection unveiled Beam, a model for coding, reasoning and AI agents. It has 501 billion parameters but uses about 23 billion at a time. This is called a mixture of experts: only the parts suited to the task are activated.
Reflection says Beam can match larger open models in several tests while using less computing power. However, these results come from the company itself. The technical report, model card and full safety evaluations have not yet been published.
The most useful thing to know now is what you can actually do. Beam is still undergoing final tests and is only available in early access to selected users. Reflection says it will release the model weights and tools under Apache 2.0 later in October.
It will not be an obvious project for an ordinary laptop either. All 501 billion parameters still need to be stored even if only some work at once, and Reflection has not yet stated practical hardware requirements. Beam has been unveiled – but you cannot freely download and run it today.