Reflection's Beam Is the New Open-Weight Model Everyone Is Talking About

| Source: Towards Data Science

Tags: Reflection, Beam, mixture of experts, open-weight models, inference hosting

Towards Data Science looks at Reflection's Beam, a 501B open-weight model that activates 23B parameters per token, and argues its hosting requirements are heavier than the active count suggests.

Details

Reflection's Beam is a new open-weight model with 501B total parameters, of which 23B are active per token. The low active count makes it look cheap to run. The Towards Data Science piece argues the real cost shows up in hosting, since all weights still need to be served even if only a fraction run per token. The feed carried only a short teaser, so the detailed hardware numbers are not available here.