Size hardware for parallel coding agents · own vs rent vs API cost over 36 months. All throughput numbers are rough FP8/vLLM-class estimates — validate on a rented node before buying.
How to read this: find your model's row, then read it as two paths. The FP8 pair (min mem + runs on) is the quality-first path; the INT4 pair is the budget path — same model, compressed to about half the memory. "Runs on" tells you the cheapest machine for each path — compare it against what you'd buy.
Columns: Parameters — total is all the knowledge the model stores (this must fit in memory); active is how much runs per word generated (this drives speed). Big models are "mixture of experts" — like a firm of 100 specialists where each question wakes only two, so a 355B model answers with 32B of effort. Min mem — memory for the model itself; workers' reading memory comes on top. Runs on — cheapest machine here that fits.
About quantization: a quantized model (INT4) is the same model compressed to fit in less memory. It is not 100% as capable as the uncompressed original — but it's close, and for most work the difference is hard to notice. It's the standard trade for running a bigger, smarter model on a smaller machine.
FP8 = near-original quality (what this calculator assumes). Min memory = model weights + overhead only; each worker's reading memory (KV cache) comes on top, so real usage needs more. "Runs on" = cheapest machine in this app that fits the weights at that quantization. Macs can hold big models in unified RAM (e.g. 512GB M3 Ultra) but are far too slow for parallel workers.
The memory rule: the model's weights plus everything the workers are reading (their context) must fit in memory at the same time. VRAM (graphics memory) is the fast lane — the model runs at full speed. Spilling over into system RAM is the slow lane: it still works, but many times slower — unusable for parallel workers. That's why "does it fit" above is a hard yes/no.
Limited by VRAM (KV cache at chosen context) or throughput, whichever binds first. Bars for tiers that can't hold the model's weights show 0.