Beam: 501B parameters, 23B active

Transcript

The original narration, preserved as published. Relative dates refer to the episode’s coverage window.

Five hundred one billion parameters. Twenty-three billion active. Reflection’s Beam is taking a different swing at the open-model race.

It’s a sparse mixture-of-experts model for coding, reasoning, and agents. The big number describes the whole model; each token activates a smaller part. That smaller active count doesn’t make this a twenty-three-billion-parameter model to store. Total weights still matter for deployment.

Reflection reports four weeks of reinforcement learning on ten thousand five hundred G B three hundred GPUs, producing more than one hundred million rollouts.

The pitch is comparable reasoning with three to four times less inference compute than G L M five point two. That’s Reflection’s benchmark comparison.

Those estimates leave out prompt processing and serving overhead. They aren’t measured hosting bills, and we haven’t independently reproduced the results.

Access is opening through a beta waitlist. The weights, model card, and technical report are promised later this month, with Apache two point oh licensing.

The next useful comparison is a complete workload: same task, same quality bar, actual latency, actual cost. That will tell us what this efficiency buys.