Culture
Decentralize Now.
A file you can run is yours. An account is a rental.
On November 13, 2026, DHH told people to keep a few open-weight models and some basic gear to run them. He did not claim to know what had happened at NVIDIA. This page does not either. Pick a kit. Mark a step when it is actually done.
A machine of your own
about $1,200–1,700
0 of 8 marked. The next unmarked step is the one that matters.
Choose the kit
Prices are street prices from November 13, 2026. Open the listing before you pay. Marks stay on this browser.
Write one sentence: what you still need if the website closes. A conversation. A page of code. A picture. A search of your own files. If you cannot name it, you are collecting, not preparing. One job is enough to start.
Weights are files. A 2TB NVMe is the working shelf: the models you actually run. On November 13, 2026 a new 2TB drive was about $250 to $300. An 8TB external hard drive, the kind that sits in a drawer, was about $250. Do not buy an 8TB NVMe. Those were $1,600 and up, and a cold copy does not need to be fast.
2TB NVMe, current listings8TB external drive, current listings
A small model runs on the computer you have. A useful one wants memory. A used NVIDIA RTX 3060 with 12GB was asking about $350 to $450 on November 13, 2026. A 16GB card was about $750. A 24GB RTX 3090 was about $2,000, which is a want, not a start. A Mac mini works too if it has 16GB of memory or more. Buy the card only if a desktop you already trust can take it.
Three programs run a model with the network off, after you have installed them. Ollama is one command and a local address other programs can call. LM Studio is a window: you see which file fits the memory you have, then you chat. llama.cpp is the engine under both, and the one to copy onto a machine that will never be online again. Install one. The program is not the model. Get it from the project, not from a file in a thread. If you use Ollama, set OLLAMA_NO_CLOUD=1 so it does not offer a cloud model when you are trying to stay on the disk.
The file name ends in the quant. That suffix is how hard the weights were squeezed. Download one file, the row that matches the memory you have. Do not keep Q4, Q5, Q6, and Q8 of the same model. The next size up replaces the last one. Q4_K_M is the default: the K_M mix keeps the important layers more exact. Q4_K_S is a little smaller and a little worse. Skip Q2 and Q3 unless nothing else fits. Skip F16. That is the unsqueezed archive, about four times Q4, and it is not the copy you run.
Which quant to download Memory Download this Size 8 GB of memory A 4B model, Q4_K_M about 3 GB 12 GB card, or 16 GB Mac An 8–14B model, Q4_K_M about 5–9 GB It loaded with room left The same model, Q5_K_M or Q6_K about 1.3–1.5× Q4 The whole file still fits, with a few GB spare The same model, Q8_0 about 2× Q4 Unplug the ethernet or turn off the wifi before you open the program. Ask a question you could check. If it answers, the file on the disk did the work. If it hangs, looks for a login, or names a cloud model, you are not offline yet. A tool that fetches the web, or a chat window pointed at someone else's server, will fail this test on purpose. Do not buy anything else until this step is done.
A backup on the same drive is not a backup. Copy the model file and the program onto the drawer drive, then unplug that drive. The people who posted their shelves said this and then admitted they had not done it. Do it before you download a second model.
Kiwix reads a ZIM file offline. The full Wikipedia is the one to get if you have the room. A smaller survival set is about 80GB. Download it, open one article with the network off, and copy the file to the drawer drive. A model that cannot look anything up is a talker. The encyclopedia is the reference.
Finished means the network is off and the model still answers, and the spare disk is unplugged. A longer list is not a higher score.
November 6 and 8, 2026
One file. No account. No card.
Ronin posted a four-step way to run a Gemma 4 file on a GPU you already have, then a short list of which file he would pull for which job. The cost he named was $0. The program is still Ollama. The model is still a file. Abliterated, the word he uses, means the refusals were cut out. It does not mean the model got smarter. It is still the size it says it is.
1. Install a current Ollama
Get it from the project, not from a command pasted out of a screenshot. An older build does not speak the Anthropic API and does not run Gemma 4.
2. Pull one named file
ollama pull igorls/gemma-4-12B-it-heretic-GGUF. He picked this one because the Heretic cut took refusals from 99 of 100 down to 0 of 100, at a KL divergence of 0.028 from the original. A random tag that only says uncensored is the mistake. The license under it is still Google’s Gemma license. Fine for personal use. Read it before you sell anything built on it.
3. Turn thinking off
Gemma 4 is a hybrid. The cut was made for the mode that does not think out loud. Leave thinking on and the reasoning block spends the token budget, then the answer stops. After the file is running, type /set nothink. If you call the API, send think set to false.
4. Point a coding tool at it, if you want
Optional. ollama launch claude --model igorls/gemma-4-12B-it-heretic-GGUF. Or set ANTHROPIC_BASE_URL to http://localhost:11434 and ANTHROPIC_AUTH_TOKEN to ollama, then name the same model. Two catches. Ollama starts with a 4,000 token window, and a coding tool’s instructions are already bigger than that, so the model never sees them. Restart the server with OLLAMA_CONTEXT_LENGTH=32768. Setting it in the shell does nothing if the server is already running. That window needs more than the 8 GB below. And Gemma 4 is clumsy with tools. For that work he names Qwen 3.5 or GLM-4.7-flash instead.
- Video memory
- 8–12 GB for the 12B at Q4.
- Disk
- About 8 GB for Q4. About 13 GB for Q8.
- A long window
- More of both. The 32k context is not free.
Which file, for which job
His list from November 8. Search the name and take the GGUF, Q4_K_M unless the table above says you have room. The pull is ollama pull hf.co, then the repo, then :Q4_K_M. Run it and type /set nothink.
| Job | File |
|---|---|
| The one he would keep | Qwen3.8-27B, abliterated |
| A small GPU | Gemma 4 12B, abliterated |
| The fast one | Gemma 4 26B-A4B, abliterated |
| A tool that edits files | Qwen3.6-35B-A3B, uncensored |
| Writing | Mistral Small 3.2, abliterated |
| No cut. As published. | Hermes 4.3-36B |
| Pictures | FLUX.1-dev, in ComfyUI |
| Video | Wan 2.2, in ComfyUI |
| Not for a house | GLM-5.1, abliterated. 754B. |
The GGUF shelf he namedHermesOllamaThe four stepsThe list
Fewer refusals is not a better mind. He still uses a cloud model for serious coding. The thing you actually keep is the weights, and a prompt that never leaves the machine. Do not use the file for anything illegal or harmful. That is on you, not on the download.






























