Where the free GPUs actually are

Four real options, and what each costs you

This lesson is deliberately specific, because “use a GPU” is useless advice to somebody who does not have one, and the free tiers are usable if you know their edges.

Google Colab, free tier. A T4 with 16 GB of GPU memory, about 12 GB of system RAM, and a session that ends after roughly a few hours: sooner if the tab is idle, and availability is not guaranteed when demand is high. Good for: fine-tuning a small model with LoRA, running anything up to about a 7B model in 4-bit, exploring a dataset. The T4 predates bfloat16 support, so use float16 and watch for the overflow failure.

Kaggle Notebooks. Around 30 hours of GPU per week, with a choice of a P100 or two T4s, and sessions up to nine or twelve hours. The weekly quota and the longer sessions make it the better free option for anything that needs to run overnight. Datasets can be attached from Kaggle or pulled from the Hub.

Hugging Face Spaces, free CPU. Two vCPUs and 16 GB of RAM, running continuously, sleeping after inactivity. No GPU. Right for embeddings, small classifiers, ONNX models and anything the previous lesson covered.

ZeroGPU on Spaces. Shared A100-class GPU time granted per function call rather than per Space, available to Spaces built with Gradio and allocated through a decorator. Free for the Space owner within quota, with a larger quota for paying accounts. It is the only way to publish a free public demo that needs a real GPU, and the constraint is per-call duration rather than total uptime, which shapes what you can build, favouring short inferences over long batch jobs.

Working with a session that will end

Everything about free compute follows from one fact: the machine is temporary. Write code that assumes it.

1. Save to the Hub, not to the session. A private model or dataset repository is durable storage that costs nothing.

trainer.push_to_hub()                  # or
api.upload_folder(folder_path="./out", repo_id="you/run-1")

1. Checkpoint often and resume. save_steps set to something you can afford to lose, and resume_from_checkpoint=True on restart.
2. Mount Drive on Colab for intermediate files if you prefer, but the Hub survives the notebook and is easier to share.
3. Do not download datasets you will stream. The disk is small and the download is repeated on every restart.

The thing that wastes most free GPU time

Installing packages. A fresh session spends several minutes on pip install before any work happens, and if you are iterating on a bug you pay it repeatedly. Two habits help: pin versions so resolution is fast and reproducible, and get the code correct on CPU with a tiny model first, then switch the model name and the device. A script debugged on a 0.5B model on CPU almost always runs unchanged on a 7B model on GPU.

Costs, when free runs out

Rental prices move constantly, so treat these as orders of magnitude rather than quotes. Community GPU marketplaces rent a 24 GB card for something in the region of a few tens of US cents per hour, and an 80 GB card for a small number of dollars per hour. Hugging Face’s own hosted options and the large clouds sit above that, buying reliability and integration.

The arithmetic that matters more than the rate: a LoRA fine-tune of a 7B model on a few thousand examples is a matter of hours, not days. At those rates it is a few dollars. People assume fine-tuning is expensive because full training of a foundation model is expensive; adapting one is not, and the training module returns to this.

The rules you must not break

Free tiers are offered on terms. Mining cryptocurrency, running a public service off a notebook, chaining accounts to evade quotas, and leaving jobs running with nobody watching are all against the terms of one or more of these providers and get accounts suspended. The quotas exist because the compute is scarce.

Do this now

Open a free Colab or Kaggle notebook and run !nvidia-smi. Write down the GPU name and the memory. Then compute, from the previous lessons, the largest model you could serve on it at bfloat16, at 8-bit and at 4-bit. That is your working envelope, and knowing it stops you starting projects that were never going to fit.