Free GPU for deep learning: 8 options and their real limits

Which free GPU should you use?

Start with Kaggle if you want the biggest weekly quota, or Google Colab if you're running quick experiments. If you need newer GPUs and don't mind spending a credit balance, look at Lightning AI and Modal. Save the big cloud trials for last.

Compare the free GPU options

Quotas and prices as of October 2026, from each provider's own pages.

The 8 options

Kaggle

Kaggle gives you 30 GPU hours a week, "or sometimes higher depending on demand and resources", and the quota resets weekly. You choose 2 NVIDIA T4s or 1 P100, each with 4 CPU cores and 29 GB of RAM, and 20 GB of output is saved in /kaggle/working. If you pay for Colab Pro or Pro+, link the accounts to get 15 or 30 extra GPU hours a week on Kaggle, a limited-time promotion.

The limit that bites is the session clock. CPU and GPU sessions run for up to 12 hours, TPU sessions for 9, and 20 GB fills up fast if you save a checkpoint every epoch. "Save & Run All" runs the whole notebook again in a new, clean session, which also has to finish within 12 hours.

Google Colab (free tier)

Colab's free tier gives you a hosted notebook with no setup. Notebooks run for at most 12 hours, "depending on availability and your usage patterns."

Google says free-tier GPU access is "heavily restricted", and it doesn't publish GPU quotas or idle timeouts. "Runtimes will time out if you are idle," and you don't know when you'll be cut off from a GPU or for how long. It suits quick experiments where a dropped runtime costs you nothing.

Lightning AI

Lightning's pricing page promises "up to 80 free GPU hours to start", while the Free plan card says up to 30 credits. You get 1 free Studio that runs around the clock but needs a restart every 4 hours, up to 2 concurrent GPUs, and single GPUs like the L40S, A100, H100 and H200.

The catch is how fast credits go on big GPUs. Lightning's own table puts the free hours at 75 on a T4, 31 on an L4, 10 on an A100 40GB, 5 on an L40S and 3 on an RTX Pro 6000, and shows none for the H100 or H200.

Amazon SageMaker Studio Lab

Studio Lab is free: you request an account, and the first time you start a runtime you verify a mobile phone number. You get 15 GB of persistent storage and 16 GB of RAM.

GPU sessions run for up to 4 hours at a time and 4 hours in any 24-hour period; CPU sessions get 4 hours at a time and 8 per 24 hours. The instance stops when you hit the limit, and compute availability isn't guaranteed. Use it if your runs finish in under 4 hours.

Modal

Modal's Starter plan has no monthly fee and includes $30 of compute a month, billed per second. GPUs range from a T4 at $0.59 an hour through an H100 at $3.95 to a B300 at $7.10, so $30 covers about 50 hours on a T4 or about 7.5 hours on an H100. Graduate students, labs and researchers can apply for up to $10k in credits.

The catch is the workflow: instead of a notebook VM, you write your training code as Modal functions and run them from your own machine. It suits developers who already work from scripts.

Paperspace Gradient

Paperspace, now part of DigitalOcean, still lists a free GPU machine (M4000) on its Gradient Free plan, alongside a free CPU machine. Free machines run notebooks in private workspaces only, and the Free plan includes 5 GB of storage. Pro costs $8 a month and Growth $39 a month, unlocking faster free machines up to an A100-80G on Growth.

The M4000 is the oldest GPU on Paperspace's price list, so it only makes sense for small models once you've used up Kaggle and Colab.

Hugging Face ZeroGPU

ZeroGPU gives Hugging Face Spaces shared NVIDIA RTX Pro 6000 Blackwell GPUs. A free account gets 5 minutes of GPU time a day, PRO at $9 a month gets 40 minutes, and the quota resets 24 hours after your first use.

Five minutes a day isn't a training budget. Use ZeroGPU to host a public demo of a model you trained elsewhere.

Cloud trial credits (Google Cloud and Azure)

New Google Cloud accounts get $300 in credit for 90 days, and a new Azure account gets a $200 credit for up to 30 days.

On Google Cloud, you can't add GPUs to VM instances while the account is on the free trial. Upgrading to a paid billing account unlocks them, and you keep the remaining credit until the 90 days run out. On Azure, 30 days goes by fast, so have your data and code ready before you sign up.

Make free hours count

Free sessions end without warning. Save checkpoints to persistent storage often, and keep your dataset small enough that a restart doesn't spend half the session on downloads.

Use the drop-in tqdm from tqdm_live in Kaggle and Colab notebooks, because the notebook widget bar isn't streamed. Kaggle notebooks have internet enabled by default; competitions that need it disabled can't send pings. Keep the stall window at 30 minutes or less.