Running Uncensored Models on a Free 32 GB Kaggle Server

The promise is simple and practical: instead of buying a bigger graphics card, you borrow a free Kaggle notebook that combines two T4 accelerators into 32 GB of memory and run any uncensored model online, then point your favorite agent harness at it as if it were a local server.

The video opens with a familiar frustration from two previous episodes about unrestricted models; viewers loved the ideas but most machines ran out of VRAM, the install failed, and the demo never started because consumer laptops rarely exceed 12 GB.

Kaggle provides that missing headroom for free: choosing GPU T4 x2 gives you two T4 cards totaling 32 GB, a weekly quota of 30 hours that resets every week, and a hard requirement to verify your phone number before any accelerator becomes available, otherwise you stay on CPU only.

After verification the flow is straightforward: hit Create to open a new notebook, open Settings and set Accelerator to GPU T4 x2, and you share the same 30 hour quota whether you pick P100 or T4 x2, with T4 x2 offering more memory and better mixed precision throughput for large models, a limit that feels more predictable than the 12 hour session cap on Colab free.

The notebook does not look like a classic server; it is a stack of code cells that only run when you prefix the command with an exclamation mark, you execute each cell individually with the Play button, and you add or remove cells with the plus control, switching between Code and Markdown as needed.

Setup begins with tiny system tools such as zstd, curl and wget installed quietly via apt, then the official Ollama install script fetched with curl and piped to sh, and because Kaggle notebooks lack systemd you must start the daemon in the background with a subprocess call or the service will block.

Once installed you run ollama serve and when the log prints Ollama started the service listens on port 11434 on the local address, which is still invisible to any external harness, so a public tunnel is required before an agent can reach it.

Model discovery happens on Hugging Face: filter for Text Generation and the Ollama app, type uncensored in the search box and the most downloaded open variants appear, letting you compare Q4_K_M and Q6_K quantized builds and see a green check when the chosen file fits the 32 GB budget, with some models offering a 16 GB VRAM optimized build.

To fetch the model you run ollama pull with an exclamation prefix inside the notebook; the progress prints line by line, the system tries to run the model once download finishes, and you can stop that attempt immediately because final inference will happen through the tunnel, not inside the notebook.

To expose the local service you install the Cloudflare tunnel binary, unpack it and launch it pointed at localhost on the Ollama port, which yields a public URL that you can open in a browser to confirm Ollama is running, and if the check fails re-running the same cell usually fixes it.

The last leg is the Hermes Agent connection: pick Custom Endpoint in the model menu and enter its number, paste the public URL with /v1 appended as the Base URL, type ollama as the API key, select the pulled model, and if the first call times out retry once after the service warms up and answers the who are you test noticeably faster than the paid cloud demo, with the full command list shared as a TXT in the creator's Discord.