Fitting the model is not feeding the model.

The saved hypothetical sample has seven hundred forty four billion parameters at four bits. Bits divided by eight gives half a decimal gigabyte per billion parameters, so disk weights total three hundred seventy two gigabytes. Resident inputs are twelve billion dense parameters giving six gigabytes, eight active two billion parameter experts giving eight, plus four KV and five buffer gigabytes: twenty three resident gigabytes. Bars share one pixel per gigabyte. This excludes real implementation overhead and does not establish a hardware benchmark, actual memory fit or source model performance. The active expert input is eight gigabytes per token before the cache assumption. At twenty five percent cache hits the missed fraction is seventy five percent and streamed expert traffic is six gigabytes per token. At fifty percent hits it is four. Bars share sixty pixels per gigabyte per token. These bytes are an analytical input workload, not measured accesses, validated cache behavior or a guarantee of prefix reuse. Dense weights, KV and other transfers are not included in this traffic formula. At the source default seven gigabytes per second and six gigabytes per token, the quotient is about one point one seven tokens per second. A target of two tokens per second requires twelve gigabytes per second. Bars share thirty five pixels per gigabyte per second and compare available seven with required twelve. This is a storage-only envelope, not end to end inference throughput: prefetch, latency, kernel overhead and compute are absent. The source explanation mentions about twenty five gigabytes even when edited resident inputs differ, so that text is not an independent fit check.