Start here

Pick the closest shape, then change anything. Every field below stays editable.

Who uses it

How many people, and how often they come back.

People using it

Sessions each, per day

Working days / month

Compare over (months)

Busiest moment Share of people typing at once. 0.05 = 1 in 20.

Reading speed (tok/s) Below ~20 an answer feels slow. This also sets how many requests one GPU can take at once.

What they do

Shares of traffic. They must add to exactly 1 — a mix that doesn't is refused, never quietly rescaled.

Chat

RAG

Graph RAG

Agentic

Tokens per turn

Workload	Turns	In	Out	Cached
Chat
RAG
Graph RAG
Agentic

The model you'd run

This is what decides how many GPUs you need. A big model, a long context or full attention can change the answer by more than 10×.

Model Pick the closest match. Every field below fills in from it and stays editable.

Size (billion params) What must sit in memory.

Active (billion) What gets read per token. Same as size unless it's a mixture-of-experts.

Context (tokens) How much history each request carries.

Weight precision Smaller = less memory, more speed.

Serving stack vLLM or SGLang if you're unsure. This only sets how much of the card's bandwidth you actually get.

Serving mode Interactive keeps every answer above your reading speed. Batch packs in more requests and accepts slower ones.

Architecture detail

KV cache precision Halving this roughly doubles how many requests fit at the same context.

Max concurrent requests Leave blank unless your runtime caps it. Memory and speed already bound this.

Attention layers The preset's own layer structure, editable. This is what decides the cache cost per token: a sliding window stops growing past its span, GDN layers never grow at all, and MLA stores one compressed latent instead of keys and values. A hybrid model keeps one block per group.

KV bytes per token override Expert shortcut. Leave blank and the layer structure above is used. Fill it in and that structure is bypassed entirely — including window-based retention, which a single flat number cannot express.

Hardware & prices

The card you'd buy, the provider you'd rent from, and the API you'd compare against.

Accelerator (owned & self-hosted) The card you'd buy. Its memory decides what fits; its bandwidth decides how fast.

GPUs

Measured tok/s per GPU Your own benchmark. Overrides the model above.

Rent from

Rented accelerator

Utilization (0–1) How busy you keep a rented card. 0.7 = idle 30% of the time and still paying.

Price feed

API model

Money

What the owned option costs you outside the tokens.

Hardware, one-time ($)

Running, per month ($)

Service level, routing & overlay

Monthly token budget

Required p95 tok/s A speed 95% of requests must beat. Left empty, no feasibility claim is made.

Price as of The instant prices are read at, declared rather than taken from the clock, so the same inputs always give the same quote.

Routing policy Which option handles a request when both could.

% self-hosted Used only by the fixed split.

Failover share Share of traffic that spills to the fallback. 0 = none.

Failover rate × How much dearer that spilled traffic is. 2 = twice the price.

Licensing ($/mo)

Consulting ($/mo)

Implementation ($) One-time, like the hardware.

Per-1M basis Whether the per-million figure carries licensing and consulting too.

Rates live only in this page — nothing is committed, published, or folded into unit prices.

Updates as you type.