Reflection AI's Beam is an open-weight model with 501 billion parameters, roughly 933 GiB of weights at 16-bit precision, a size that matches the published BF16 checkpoint on Hugging Face. The headline figure most people repeat is 23 billion active parameters per token. That number is real and useful, but it does not reduce the memory needed to store the model.
Beam is a sparse mixture-of-experts model. A router selects a few expert networks for each token, so only those experts do work on that token. Active parameters determine the computation and price per token, but every expert must stay in memory because the router can send the next token to a different set of experts. The cost shifts from compute to memory rather than disappearing, and memory is the cost you pay before the first token is generated.
Reflection says Beam matches GLM-5.2 on advanced reasoning benchmarks while using three to four times less inference compute. That claim comes from Reflection, and Tech Crunch notes that the performance and efficiency claims remain independently unverified. Reflection also acknowledges that Kimi K3 is ahead on raw capability, so the pitch rests on efficiency. An estimate of compute based on active parameters and reasoning tokens does not capture measured latency, throughput, energy use, or provider pricing, which depend on hardware and software.
Settings can also change costs. Beam's API defaults to medium reasoning effort, while OpenRouter lists max as its route default, so the same prompt can cost very different amounts depending on where it runs.
Source: Towards Data Science · Summarized by HeadlinesBriefing