Reflection has launched the Beam model with 501 billion parameters in a Mixture-of-Experts (MoE) architecture. It is optimized for efficient inference on coding and agent-based workloads.
Reflection has released Beam, the first model in its new family, using an MoE architecture with 501 billion total parameters, 23 billion active per token, and a context window of up to one million tokens.
Return to original text
A short introduction is retained here; the full text from the source website is not reposted. For specific wording and subsequent updates, please refer to the original article.
Open AIHOT original