Kaggle Notebooks (free GPU/TPU hours)

30 free GPU hours per week (choice of dual Tesla T4 or P100), plus 20 free TPU hours per week; the quota resets weekly, and a single session runs for up to 12 hours

This is a compute quota, not a token/request quota — users load their own models inside a notebook to run inference or training. GPU options include a single T4, dual T4 x2 (32GB VRAM combined), or the older P100 (16GB). Also includes a 20GB per-session working disk and unlimited access to public notebooks, datasets, and competitions. Compared to Colab, Kaggle's weekly quota figures are clearer and community feedback more stable, with no major cutbacks recorded during 2026. Requires phone verification to enable GPU and internet access.

⚠️Also owned by Google, but with far more transparent and stable quota rules than Colab (a fixed '30 hours/week' rather than Colab's dynamic fluctuation) — dual T4 x2 with 32GB combined VRAM is enough for individuals to run inference or fine-tuning experiments on quite a few 7B-13B-class models. The downside is that it's limited to the notebook environment and can't be called as a regular API, and the phone-verification bar is slightly higher than Colab's.