Beam VRAM calculator

Beam has 501B total parameters. Only 23B are active per token, which makes it fast, but every expert still has to sit in memory. Pick a format and context length to estimate the footprint.

Weight format

Context length (tokens)

KV cache per 1K tokens (GB)

Weights

–

KV cache

–

Runtime overhead (~5%)

–

Estimated total

–

Devices needed

Device	Count

The KV cache figure is an assumption until Reflection AI publishes the model config (layer count, attention type). It is editable above. Device counts assume about 90% of each device’s memory is usable and ignore interconnect limits.