SARCASM WARNING! (Also: code warning. There’s actual working stuff buried in here, so don’t skim too hard.)<\/p>\n\n\n\n
Well hello there, my GPU-starved friends! Tired of watching your laptop wheeze like an asthmatic hamster every time you try to run a real AI model? Good news. Kaggle will hand you two 16 GB GPUs, about 30 hours a week, for the low, low price of absolutely nothing.<\/p>\n\n\n\n
Here’s the math. A 27-billion-parameter language model needs about 17 GB of GPU memory just to wake up. Your laptop has 8 GB of RAM, a sticker from a 2019 conference, and dreams. Kaggle has the hardware. You have the audacity. Let’s make a deal.<\/p>\n\n\n\n
Here’s the whole trick: take a Kaggle notebook (built for data science homework), turn it into a private AI server, load an “uncensored” Qwen model onto it, punch a hole to the internet with a Cloudflare tunnel, and point Claude Code at it from your own machine. Fifteen minutes. Zero dollars. What could possibly go wrong?<\/p>\n\n\n\n
Spoiler: one thing. One very large, very open, very front-door-shaped thing. I went through every command so you don’t have to, fixed the ones that quietly break, and found the problem nobody seems to mention. It’s the most important part of this post, so naturally I’ve put it near the end, where nobody reads.<\/p>\n\n\n\n
Kaggle exists so people can enter machine-learning competitions and argue about leaderboards. To keep that fair, it gives every verified account free notebook time on real GPUs. Real ones. Not the “integrated graphics” kind your IT department calls a workstation.<\/p>\n\n\n\n
The option you want is GPU T4 x2<\/strong>: two NVIDIA T4 cards with about 15 GB of usable memory each.<\/p>\n\n\n\n Why two? Because the model, a 4-bit build of Qwen3.8-27B, weighs roughly 17 GB, and one T4 can’t hold it. Kaggle’s other option, a single P100, can’t either once you leave room for the actual conversation. Two T4s can, and Ollama splits the model across them without being asked. Teamwork. Something your SOC and dev team could learn from.<\/p>\n\n\n\n Before anything works, flip two switches:<\/p>\n\n\n\n Now the fine print, because there’s always fine print. Sessions die after about 12 hours or when idle. The weekly quota is about 30 GPU hours. And every new session starts from absolute zero: a fresh 17 GB download and a brand-new public URL. It’s Groundhog Day, but with CUDA. Remember this. It comes back to bite later.<\/p>\n\n\n\n The heart of all this is Ollama, a tool that downloads open models and serves them over a local API. Getting it running on Kaggle takes four steps. Three of them are booby-trapped, and the obvious way of doing it strolls right over them like a tourist in a minefield. Here’s the damage report:<\/p>\n\n\n\n Trap one: the installer needs a tool Kaggle doesn’t have.<\/strong> Ollama ships as a The installer will then complain that systemd isn’t running. Ignore it. Notebooks have no service manager, so you start the server yourself, like an animal.<\/p>\n\n\n\n Trap two: the default memory is goldfish-sized.<\/strong> Ollama gives every model a 4,096-token context window unless told otherwise. Sounds like plenty, until you learn that Claude Code’s opening instructions alone are longer. And here’s the kicker: Ollama doesn’t throw an error when a prompt overflows. It quietly chops off the beginning and carries on, smiling. Your agent forgets what it was doing and you never find out why. It’s walking into a room and forgetting why you came in, except it happens on every single message. So start the server with a bigger window:<\/p>\n\n\n\n And yes, Trap three: That odd-looking name is Ollama pulling a GGUF file straight from Hugging Face. Before anyone gets excited: “uncensored” does not mean smarter, edgier, or secretly aware of where the bodies are buried. The model was made with a technique called abliteration, using a tool named Heretic (yes, really). It finds the direction inside the model that produces refusals and removes it, with no retraining. On the author’s 100-prompt test of harmful requests, refusals fell from 98 to 12.<\/p>\n\n\n\n That’s it. That’s the whole upgrade. It is not smarter, and removing pieces of a model usually costs it a little accuracy. It’s like taking the brakes off a car and calling it a sports car. If you want a coding assistant and don’t care about refusals, the stock Qwen model is the better pick.<\/p>\n\n\n\n Congratulations, you now have a 27B model running on Enter the Cloudflare quick tunnel, which solves this by going the other way. The \n
\n\n\n\n\ud83e\udea4 Four Commands, Three Traps<\/h2>\n\n\n\n
Step<\/th> The obvious way<\/th> What actually happens<\/th> Sanity level<\/th><\/tr><\/thead> Install Ollama<\/td> Runs the official install script<\/td> Fails: Kaggle has no zstd<\/code><\/td>Mildly shaken<\/td><\/tr> Start Ollama<\/td> Trusts the defaults<\/td> 4,096-token goldfish memory<\/td> Suspicious<\/td><\/tr> Get the model<\/td> ollama run<\/code><\/td>Cell waits forever for someone to type<\/td> Philosophically broken<\/td><\/tr> Open the tunnel<\/td> Reads 30 lines of log, then stops<\/td> Tunnel can freeze hours later<\/td> Who am I anymore?<\/td><\/tr> Secure it<\/td> API key: ollama<\/code><\/td>Protects absolutely nothing<\/td> Numb but enlightened<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n .tar.zst<\/code> archive, and Kaggle’s image has no zstd<\/code> to unpack it. That’s IKEA delivering your wardrobe without the Allen key. Install it first:<\/p>\n\n\n\n!sudo apt-get install -y zstd\n!curl -fsSL https:\/\/ollama.com\/install.sh | sh<\/code><\/pre>\n\n\n\nimport os, subprocess, time\n\nenv = os.environ.copy()\nenv[\"OLLAMA_CONTEXT_LENGTH\"] = \"32768\"\nenv[\"OLLAMA_KEEP_ALIVE\"] = \"-1\"\nsubprocess.Popen([\"ollama\", \"serve\"], env=env,\n stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)\ntime.sleep(5)<\/code><\/pre>\n\n\n\nPopen<\/code> matters. Run !ollama serve<\/code> directly and the cell blocks forever, just sitting there. Staring. Judging.<\/p>\n\n\n\nrun<\/code> waits for a person who isn’t there.<\/strong> The obvious way to download the model is ollama run<\/code>. That command downloads, then opens an interactive chat and waits for you to type. A notebook cell has no keyboard. So it waits. And waits. Like a Tamagotchi nobody’s feeding. Use pull<\/code>, which downloads and actually leaves:<\/p>\n\n\n\n!ollama pull hf.co\/JonathanColetti\/Qwen3.8-27B-Uncensored-GGUF:Q4_K_M<\/code><\/pre>\n\n\n\nQ4_K_M<\/code> is the 4-bit version, the usual trade-off between size and quality. Think of it as the model’s travel-size shampoo.<\/p>\n\n\n\n\ud83d\ude08 What “Uncensored” Actually Means (Calm Down)<\/h3>\n\n\n\n
\n\n\n\n\ud83d\ude87 A Tunnel Out of a Building With No Doors<\/h2>\n\n\n\n
127.0.0.1:11434<\/code>, inside a container, inside a Google data centre. Nothing outside can reach it. Kaggle notebooks accept no incoming connections. Your shiny new AI server is basically a genius locked in a basement.<\/p>\n\n\n\ncloudflared<\/code> program inside the notebook dials out<\/strong> to Cloudflare and holds the line open. Cloudflare hands you a public address like https:\/\/some-random-words.trycloudflare.com<\/code>, and requests to that address travel down the line to Ollama. No account, no domain, no firewall rules, no change request, no CAB meeting. Your GRC team would faint.<\/p>\n\n\n\n