TIA · HYDRA · SORTIE · AI SIZING

Which model. Which hardware.
Measured, not estimated.

You want to run AI in-house but don't know which model to pick, how much memory it will need, or whether it can handle your task in your language at all. Instead of guessing, we measure it. We run candidate models on a rented GPU against tasks that look like yours, and within one working day you get numbers. Then the machine is destroyed.

Your real data never goes to the rented machine.
We measure on synthetic, public, or your own anonymised samples.
1 dayto results
11 minpilot, 3 models
0of your data on the GPU
from CZK 5,000entry measurement
WHAT SORTIE ANSWERS

Three questions that usually get answered with a guess.

01
Which model
Quality on your task
  • Measured on your task, not a generic leaderboard
  • Including what leaderboards miss: whether the model correctly refuses when the answer isn't in the documents
02
Which hardware
Memory and speed, measured
  • Actual GPU memory used, in GB
  • Throughput in tokens per second
  • Server sizing to the gigabyte, not to a vendor's estimate
03
What isn't worth it
Often the most valuable line
  • A bigger model that adds nothing on your task except twice the hardware cost
  • You can't tell by estimating — only by measuring
HOW IT WORKS

Four steps. One working day.

1 · Brief

You describe the task (contracts, e-mail, documentation, support…) and the constraints: on-prem, language, budget, data sensitivity.

2 · Sample

We build representative tasks: synthetic, public, or from a sample you anonymise. Your real documents stay with you.

3 · Measurement

Candidate models run on a rented GPU (80 GB). The run takes minutes; afterwards the machine is destroyed automatically — even if anything fails.

4 · Result

A table, a recommendation, and a measurement record you can check afterwards: what was measured, on what, with what result.

EXAMPLE · PILOT FOR A LAW FIRM · ANONYMISED

Which open model can handle Czech legal documents?

Seven tasks: extracting parties, deadlines and penalties from contracts, summarising a judgment, working across several documents, a statute question — and one task where the correct answer is “it isn't in the documents”. All data was synthetic.

ModelGPU memorySpeedTasks passed
mistral-small 24B19.4 GB30.7 tok/s6 / 7
gemma3 27B28.3 GB15.1 tok/s7 / 7
qwen2.5 72B55.8 GB15.4 tok/s7 / 7
What followed: the largest model (72B) added nothing on these tasks for roughly twice the memory. The 27B model passed everything and fits a 32 GB card, with a healthy margin up to 48 GB. The fastest, 24B, failed one task out of seven. Measuring all three models took 11 minutes on a single A100 80 GB; compute cost under one dollar.
PRICING

A small entry fee. Anything larger is priced up front.

ENTRY
Entry measurement
A fast preliminary result
from CZK 5,000
/ €200 · excl. VAT
⏱ within 1 working day
  • One to three candidate models
  • A task set built from your brief
  • Memory, speed, quality, recommendation
+ EXTENDED
Extended measurement
Priced by models and variants
custom
/ always confirmed in advance
⏱ depends on scope
  • More models and more configuration variants
  • A larger task set built with you
  • A measurement record for later review

What needs saying before anyone asks

  • It is not a production load test with hundreds of concurrent users; we measure the model and the task, not your whole application.
  • Results hold for the measured configuration and task sample. Seven pilot tasks is a small sample; for an engagement the set is built with you and is usually larger.
  • We never measure your real data on the rented machine. Measuring on live data is only possible on your own hardware.
  • We don't sell hardware or model licences — so we have no reason to recommend anything bigger than you need.
GOT A CONCRETE TASK?

Tell us what the AI should do and where it should run.

We reply the same working day with a proposed task set and a price.

Send an e-mail
hydra@tia-framework.com · WhatsApp