Running all of this without money

What is free, and what each is good for

Nothing in this course requires you to buy hardware. Here is the honest state of the free tiers, with their real limits rather than their advertised ones.

Kaggle Notebooks is the most generous and least known. Around 30 GPU-hours per week, on a P100 or a pair of T4s, with sessions up to about 12 hours, plus a large CPU allowance. That is enough to fine-tune a 7B model with QLoRA, run a full evaluation, or process a substantial batch job. If you take one thing from this lesson, it is that this exists.

Google Colab free tier gives you a T4 with 16 GB of VRAM, sessions that disconnect after a few hours and sooner if idle, and no guarantee of a GPU at busy times. Good for a notebook you can restart, poor for anything that must complete.

Hugging Face Spaces hosts free CPU Spaces indefinitely. More usefully, you can use other people’s public Spaces at no cost, which is how to try six image or speech models in twenty minutes without installing anything.

Oracle Cloud’s always-free tier provides ARM instances with a meaningful amount of RAM and no time limit. There is no GPU, so this is CPU inference, but a 7B or 8B model at Q4 runs on it at a few tokens per second, permanently, for nothing. For a background job or a low-traffic personal service that is a real deployment.

Provider free credit. Several inference providers give a small amount of free usage, and Hugging Face includes a monthly credit for its inference partners. Enough to run an evaluation across several models; not enough to serve anything.

Your own machine. An ordinary laptop runs a 3B model at readable speed on CPU alone. This is not a consolation prize; most of the extraction, classification and rewriting work in this course runs there.

What fits where

On a T4 with 16 GB of VRAM:

• Inference on a 7B or 8B model at 4-bit, with room for a long context.
• Inference on a 13B at 4-bit, tight.
• QLoRA fine-tuning of a 7B with a modest sequence length and gradient checkpointing.
• Not: full-precision fine-tuning of anything interesting, or a 70B in any form.

On CPU with 8 to 16 GB of RAM: 3B to 8B at Q4, at the speeds the bandwidth arithmetic predicts.

Working within session limits

The constraint that actually hurts is not memory, it is that free sessions end. Three habits make it survivable, and they are the same habits that make any long job survivable.

Checkpoint to durable storage every N steps. Not to the session’s local disk, which disappears. Push adapters to a private repository, or write to mounted cloud storage.

from transformers import TrainingArguments
args = TrainingArguments(
    output_dir="out", save_steps=100, save_total_limit=2,
    push_to_hub=True, hub_model_id="you/my-adapter", hub_private_repo=True,
)

Make the job resumable. A training run that can start from the last checkpoint turns a disconnection into a delay. A run that cannot turns it into a loss.

Write results as you go. One JSONL line per completed item, appended immediately. If the session dies at item 380 of 500, you have 380 results, not none.

These are the same disciplines that matter on paid infrastructure; the free tier just teaches them faster.

What the free tiers are not for

Serving a product. Rate limits, unpredictable availability and terms that prohibit it make this a bad idea, and it will fail at the moment anyone actually uses your thing. For serving, either rent per token, which starts at a few cents, or use the always-free CPU instance and accept its speed.

The order that wastes least: evaluate on free provider credit, fine-tune on Kaggle’s weekly hours, and only then decide what to pay for. Most people buy hardware first and discover afterwards that a 3B model on their existing laptop did the job.