Free GPU: three legitimate ways

Train a model or run inference without your own graphics card. Each option has its own limits — rotate them.

1. Google Colab

• a GPU (usually a T4) in a browser notebook, nothing to install;
• sessions of a few hours, a floating daily cap;
• the best fit for study projects and prototypes;
• card: in the catalog.

2. Kaggle Notebooks

• ~30 GPU hours weekly + 20 TPU hours — the most predictable cap;
• datasets and competitions right next door;
• less internet access than Colab — bring your own datasets;
• card: in the catalog.

3. Lightning AI Studio

• full cloud machines with GPUs and a monthly credit;
• studios survive closing the tab — good for long experiments;
• card: in the catalog.

Survival rules

• checkpoint to Google Drive/Kaggle Datasets every 15–30 minutes;
• don't request a GPU when the CPU would do — caps get wasted;
• rotate platforms: hit Colab's daily cap, switch to Kaggle;
• small models are fine on free LLM APIs — 54 no-card options.

Snapshot 05.11.2026 · limits drift — check the official pages