Skip to content
Fathom Layer

Fathom Layer / Decision instruments

AI calculators.
Run the numbers.

Estimate model memory, cloud versus local costs, prompt spending and vector storage. Change the inputs and see where your decision changes.

Open the workbench Choose a question

Estimates, not quotes. Every input stays editable.

01 / Find your question

Choose the constraint.

Each instrument solves a different part of a technology decision. Start with the uncertainty in front of you.

01 / INSTRUMENTWill it fit?Model memoryEstimate weights, context cache and runtime overhead for local inference.Open instrument 02 / INSTRUMENTWhen does it pay back?Cloud or localChange workload and operating assumptions to explore a break-even point.Open instrument 03 / INSTRUMENTWhat happens at scale?Prompt volumeTurn a prompt and request volume into a comparable cost estimate.Open instrument 04 / INSTRUMENTHow much memory?Vector storageSize raw embeddings, index overhead and a working memory buffer.Open instrument

02 / Workbench

Change an input.
Watch the answer move.

Use a starting scenario or build your own. Results update as you adjust the controls; the underlying assumptions remain visible.

Instrument 01 / 04

Will it fit?

Modeled estimate
Explore the sensitivity

Keep a reference. Change one input.

Hold the current estimate, then adjust the controls. Compare how the modeled result changes; the reference stays separate for each instrument.

Current / Live117.87GiB

Modeled memory

Binary GiB footprint · weights, cache and runtime

32,768
1
CustomGiB available after other processes

Estimated memory footprint

•Weights: 36.7 GiB
•KV-Cache: 80.0 GiB
•Runtime allowance: 1.2 GiB
Total: 117.9 GiB
Your 24 GiB scenario

Weights exceed available memory

The estimated weights alone are larger than the memory you entered. Try a smaller model, stronger quantization, or more memory.

How to read this estimate

The cache calculation uses every attention head, so it is a conservative upper bound for models with grouped-query attention. Quantization format, runtime and batch behavior can change the actual requirement. Verify the model configuration and available memory before choosing hardware.

Explore hardware on record →

Results depend on the entered assumptions. Recorded prices and modeled memory are not live quotes or measured deployment results.

Continue the decision

Keep your inputs in view as you inspect the next part of the decision.

The general calculator uses decimal GB and a different cache formula; KV precision and parallel streams are not carried over.

Explore this memory scenarioExplore compute hardware

03 / Starting points

Begin with a real question.

These are editable example inputs. They are not measurements of a deployed system or current vendor quotes.

ALocal inferenceA 70B model at Q4Start with model weights and a 32K context window. Then change concurrency and see the memory move.Size the model BOperating costA heavier API workloadBegin with 50 million monthly input tokens and a hardware budget. Edit the assumptions before interpreting payback.Explore break-even CVector infrastructureTen million embeddingsExplore memory for 1536-dimensional vectors with an HNSW index. The result is an estimate, not a deployment quote.Size the index

04 / Read the result

A number is only as useful
as its assumptions.

These calculators model a scenario. Memory estimates include model weights, context cache and overhead; cost estimates depend on the workload and rate values shown in the tool. Actual hardware behavior, provider terms and deployment costs may differ.

How Fathom handles evidence Explore hardware on record