Learn
See
Try
Make

Run a model on a free machine

What this lesson is about

Load a small model in a free notebook, send it ten inputs, and save what came back.

Ten customer messages in Hindi are waiting and a free machine is open for an hour. What do you do with them?

Why this matters

One line names the model and sets it up to run, and the next line sends it your text. On a free machine the limit you meet first is memory, not patience. A free Colab notebook offers a T4 card with 16 GB, read on 14 Sep 2026. Colab says the card is not guaranteed and a session ends after at most 12 hours. Llama 3.2 1B needs 2,358 MB at full width, so it fits as it is. A model that does not fit gets quantised — stored with shorter numbers. The Transformers docs say 8-bit storage halves the memory and 4-bit cuts it to a quarter. Meta's card lists a 4-bit QLoRA build of the same 1B model at 1,127 MB. The docs' honest cost: shorter numbers keep as much accuracy as they can, not all of it. So measure again after quantising.

You can load a small model on a free machine, run ten inputs, and save what came back.

Load a small model, send it inputs, and save the outputs where you can read them.
Memory runs out before your time does; a smaller model is the usual fix.
Save the exact model name and version with the outputs, or nothing repeats.

Task

Running ten Hindi customer messages through a small open model on a free machine.

Weak approach Running all ten, reading the replies on the screen, and agreeing they look right. The tab is closed at the end of the hour and nothing of the run is left.

Stronger approach Loading Llama 3.2 1B Instruct on a free Colab T4, in full width because 2,358 MB fits in 16 GB. Sending all ten messages and putting each reply into a sheet beside its message. The sheet's top row carries the model's full name, the width it was loaded at, and the day it ran.

Why the stronger approach works A run that was not saved cannot be set beside tomorrow's run. That comparison is why you ran it.

Try a changed situation Your phone stops while you are editing a long video for a client, and the app shuts without warning. What do you try first?

Make something yourself Save a notebook that loads one model and runs ten of your own inputs. Write the outputs to a file, with the model name and the date at the top.

Lesson 3 of 8 on Open models on a small budgetnextMeasure it before you change itPart 1 ends in a work sample you can send

Read the primary or official source: Hugging Face Transformers · quickstart