Colab GPU Limits Explained: Free vs Paid, GPU Types and Setup
Google Colab gives you a notebook with a GPU in your browser, free to start. The catch is that Colab does not publish fixed free limits: how long you can run and which GPU you get changes with demand. Paid plans use compute units. This guide shows how to read the limits, pick a GPU, set up a notebook and stop wasting units.
Colab is a good place to learn and test media and AI experiments. Check every number below against Google’s own pages on the day you start, because they change.
What are the Colab GPU limits on the free tier?
Colab does not publish its free GPU limits. Google’s FAQ says usage limits, idle timeouts, maximum runtime and available GPU types all vary over time. The FAQ does say a free notebook can run for at most 12 hours, depending on availability and your usage.
So you cannot plan around a fixed number of free GPU hours. Plan around the idea that a session can end early. Save work often, write outputs to Google Drive, and keep each job small enough to restart.
Read the official FAQ yourself: https://research.google.com/colaboratory/faq.html
How do paid plans and compute units work?
Paid plans give you a compute unit balance. Units are spent while a runtime is connected, and faster GPUs spend them faster. When the balance hits zero you fall back to free-tier rules. Pay As You Go, Pro and Pro+ all raise your availability based on that balance. Pro+ supports runs of up to 24 hours when you have enough units.
Plan names, prices and unit amounts change, and the pricing page loads for signed-in users, so we are not quoting figures. Check the current numbers here: https://colab.research.google.com/signup
One number from our own testing, labelled as such: in our test (between 28 Oct and 7 Nov 2026), a paid L4 GPU used about 1.9 compute units per hour. Your rate depends on the GPU you pick and can differ, so read the rate shown in Colab before you connect.
Which GPU types does Colab offer?
It depends on the day and your plan. The free tier has typically offered a T4, a 15 GB GPU that is enough for learning and many small experiments. Paid plans can offer faster GPUs, subject to availability. Google says GPU and TPU types vary over time, so the runtime menu is the source of truth.
| Choice | Where you see it | Good for |
|---|---|---|
| T4 | Runtime > Change runtime type | Learning, small models, quick tests |
| Faster paid GPU, if offered | Same menu, paid plans | Bigger models, longer runs |
| CPU only | Same menu | Downloading, unzipping, copying files |
Pick the cheapest option that fits the job. Many steps, such as downloading files, do not need a GPU at all.
How do I set up a notebook with a GPU?
Open a new notebook, then choose Runtime > Change runtime type, select a GPU and pick T4 on the free tier. Save, then connect. Run a one-line check to confirm the GPU is attached before you do anything else:
!nvidia-smiIf you see a table with a GPU name and memory, you are set. If not, the runtime menu probably offered no GPU at that moment; try again later or switch to CPU for the non-GPU steps.
A notebook written by someone else shows a "Run anyway" warning on first run. Read the code before you click it. Only run notebooks you trust.
How do I upload and download files?
For small files, open the Files pane on the left and drag files in. They live in the runtime and disappear when it is deleted. To get a result out, use the download helper:
import google.colab.files as cf
cf.download("result.mp4")For anything you want to keep, mount Google Drive and write there:
import google.colab.drive as gd
gd.mount("/content/drive")One tip from our test: do not paste large base64 images into a cell. It froze the page for us. Upload the file through the Files pane or Drive instead.
How do I cache big files in Drive?
Download big files once, store them in Drive, and reuse them. In our test, a free T4 notebook for a video model spent about 8 minutes on setup, mostly pulling roughly 30 GB of model files, and GPU memory peaked near 14 of 15 GB. That was one run on one date, so treat it as a rough picture, not a promise.
The saving is simple: run the download step on a CPU-only runtime, copy the files to Drive, then start the GPU runtime and read from Drive. You stop paying GPU rates for waiting on a download. Check your Drive storage first, since big files fill the free quota fast.
How do I stop wasting compute units?
Disconnect when you are done. A paid GPU keeps spending units while it sits idle, so release it the moment your job ends. Use Runtime > Disconnect and delete runtime. You can also do it from code at the end of a notebook:
import google.colab.runtime as rt
rt.unassign()A few habits that help:
- Test on a tiny input before a long run.
- Do downloads and file prep on CPU.
- Write results to Drive as you go, so a dropped session costs you little.
- Do not leave a tab open on a paid GPU overnight.
What is not allowed on the free tier?
Google’s FAQ lists activities that can be stopped without warning. On the free tier these include remote control such as SSH shells and remote desktops, bypassing the notebook interface to work mainly through a web UI, and running distributed computing workers. On all managed runtimes the FAQ also bars torrenting, connecting to remote proxies, mining cryptocurrency, password cracking, creating deepfakes and using multiple accounts to get around restrictions.
Treat Colab as a notebook for interactive learning and testing. Read the full list before you plan anything unusual: https://research.google.com/colaboratory/faq.html
Colab or Kaggle: which should a beginner try?
Both give a free GPU notebook for learning. They differ in how limits are shown. Kaggle shows a weekly GPU quota on its settings page. Colab does not publish fixed free limits.
| Question | Colab | Kaggle |
|---|---|---|
| Are free GPU limits published? | No, they vary | A weekly quota is shown in Settings |
| Typical free GPU | A T4, subject to availability | T4 x2, per its docs |
| Paid option | Compute units | Not the main route |
| Files that persist | Google Drive | Saved notebook output and datasets |
For the Kaggle side, including its terms, read our guide: Kaggle GPU free limit guide. Whichever you pick, confirm the current limits on the official page first.
If you pay for Colab Pro or Pro+, Kaggle has offered extra weekly GPU hours when you link the two accounts. Check the current offer on Kaggle’s settings or docs page before you count on it.
FAQ
Q: Is the Colab GPU really free?
You can start on the free tier with a GPU when one is available. Limits are not published and change over time.
Q: How many hours of free GPU do I get?
Google does not publish a fixed number. The FAQ states a 12-hour maximum per notebook run on the free tier, depending on availability.
Q: Which GPU should I choose?
Start with a T4. Move to a faster paid GPU only when a job truly needs more memory or speed.
Q: Do compute units run out?
Yes. When your balance reaches zero you return to free-tier rules until you add more.
Q: Can I use Colab for a client project?
Read Google’s current terms and FAQ for your case before any commercial use.
Send a product image
Need a finished product video without running a GPU yourself? Send us one product image and we will return a free 15-second sample, no obligation. Agencies can see how white-label work runs on our for agencies page.
Or read how it works on the agencies page.