Google Colab (free GPU)

Free access to an NVIDIA T4 GPU (16GB VRAM); the vendor doesn't publish a fixed weekly quota — allocation is dynamic. A single session runs for up to 12 hours

This is a compute quota, not a token/request quota — users load their own models (via Hugging Face, PyTorch, etc.) inside a notebook to run inference or training. The vendor explicitly states that the available accelerator type, idle timeout, and maximum runtime vary dynamically with 'demand and account usage history,' with no guarantee a T4 will always be available. A third-party measurement sample (multi-account observation, 2026-03 through 04) shows a median weekly availability of roughly 22.4 hours, ranging roughly 15-30 hours/week depending on region and account history.

⚠️The developer community has long reported that free-tier GPU availability keeps 'getting tighter' (especially during peak hours, when it's prone to being downgraded to CPU or not allocated a GPU at all) — don't rely on its stability for serious projects. Its advantage is zero barrier to entry (just a Google account) and the most mature ecosystem, well-suited to getting started or running one-off experiments, but not for long-running tasks that need stable compute.