Skip to content

Cookbook

Can I run this AI model?

A plain answer to a confusing question. Tell it the graphics card you own, or the model you want to run, and it shows what fits, how fast it will be, and what it costs. If your card is too small, it names the cheapest one to rent. 57 models, no account, no jargon you have to already know.

  1. 1

    Pick your card, or a model

    Choose the GPU in your machine, or the model you want to run. That is the only input.

  2. 2

    See what fits, and how fast

    The table shows what runs, the memory it needs, and roughly its speed. Green means it fits.

  3. 3

    Run it, or rent a card

    Runs on your card? Done. Too big? It names the cheapest card to rent that holds it, per hour.

cookbook
Kracht

Pick the graphics card in your machine. The table lists every model, the easiest to run at the top. A green Runs badge means it fits your card; anything too big shows the cheapest card to rent instead. Precision and context change how much memory a model needs.

Reading forGeForce RTX 409024 GB usable1008 GB/sDesktop graphics card
FitScoreModelSizePrecisionMemoryContextSpeedDo it
Runs77/100Qwen3 32BChat32Bint421 GB8k~44 t/son your card
Runs77/100QwQ 32BChat32Bint421 GB8k~44 t/son your card
Runs77/100Qwen2.5 Coder 32BCode32Bint421 GB8k~44 t/son your card
Runs77/100Qwen2.5 32BChat32Bint421 GB8k~44 t/son your card
Runs80/100Qwen3 30B-A3BChatMoE30.5B /3.3Bint420 GB8k~428 t/son your card
Runs85/100Gemma 3 27BChat27Bint418 GB8k~52 t/son your card
Runs85/100Gemma 2 27BChat27Bint418 GB8k~52 t/son your card
Runs90/100Mistral Small 3.1 24BChat24Bint416 GB8k~59 t/son your card
Runs94/100Codestral 22BCode22Bint415 GB8k~64 t/son your card
Runs100/100StarCoder2 15BCode15Bint410.0 GB8k~94 t/son your card
Runs100/100Phi-4 14BChat14Bint49.3 GB8k~101 t/son your card
Runs100/100Qwen2.5 14BChat14Bint49.3 GB8k~101 t/son your card
1–12 of 57 models

57 models · 41 run on a GeForce RTX 4090 at int4, 8k context

estimates, err upward · not benchmarks

Runs now

Already own a card? Check it in one line

A web page cannot look inside your machine, and this one does not pretend to. The command below is meant to run on the box you are asking about, print what it finds, and send nothing anywhere.

curl -fsSL https://kracht.ai/diagnose.sh | sh

It reads the GPU model, its VRAM and the driver: the same three facts the node agent already reports on every heartbeat once a machine is registered. This is a one-off look for someone deciding whether to bother.

Nothing leaves the machine. There is no curl, wget or nc anywhere in it, no telemetry and no identifier, and the build refuses to publish it if that stops being true. Read it first if you would rather: kracht.ai/diagnose.sh is the same file the command fetches.

Everything here is arithmetic from a spec sheet, before a rental exists. When one does, Autopilot reads what it actually cost and how long it actually took, and says whether a different card would have been cheaper.