Free 32 gb vram for local ai models — no need to buy a graphics card

[jacob266]

Found a working way to run pretty heavy models (up to 27–32b in quant) without owning a powerful gpu.
we use free kaggle gpu notebooks: two t4s with 16 gb each = 32 gb vram. weekly limit is about 30 hours.
quick steps:

1. sign up on kaggle and verify your phone (otherwise you won’t get the dual t4).
2. create a notebook → settings → accelerator → gpu t4 × 2.
3. install ollama and start the server right in the notebook.
4. pull the model you need (from hugging face, including uncensored ones).
5. spin up a temporary cloudflare tunnel with cloudflared and get a public url.
6. on your local machine connect to that url as a regular ollama endpoint (terminal, cline, open webui, etc.).

the model runs in the cloud, you talk to it almost like a local one. the session only lives while the notebook is active, so it’s not great for 24/7 use, but perfect for testing, coding and experiments.
if anyone needs the ready-to-paste commands, just ask.