Reflection releases Beam, an open-weight model with 501 billion total parameters and 23 billion active parameters. The company says it matches GLM-5.2 on advanced reasoning benchmarks while using only one-third to one-quarter of its competitors’ compute for inference.
These figures come from the company’s own benchmarks and have not been independently verified. The model uses a mixture-of-experts architecture: it has many parameters overall, but activates only a small subset for each inference request, making each call more efficient. Note that the compute savings refer to inference; training involved compute-intensive reinforcement learning.
These figures come from the company’s own benchmarks and have not been independently verified. The model uses a mixture-of-experts architecture: it has many parameters overall, but activates only a small subset for each inference request, making each call more efficient. Note that the compute savings refer to inference; training involved compute-intensive reinforcement learning.