A bunker of free AI models on your own drive
Six free models and apps, plus Ollama to run them, with the exact commands: Qwen3.5, gpt-oss, GLM-5.3, Jan, AnythingLLM, and Ollama inside Claude’s desktop app. All but GLM-5.3 run on your own computer; its easy route is Ollama’s cloud. A memory guide, a backup kit, and what the headlines behind it actually were. Checked against the makers’ pages, the licences and Ollama’s own docs
A model on your own drive can’t be taken back. This is the short list of free ones worth having, with the exact command for each: install Ollama, pick the model your memory can hold, add Jan for a ChatGPT-style window and AnythingLLM to ask questions of your own files, and, on a Mac, run them inside Claude’s desktop app. The checks kept every command and fixed the story around them. The model “they just pulled” was never pulled: Hugging Face disabled one unofficial copy of GLM-5.3 with its safety refusals removed, and Z.ai’s own release is still up. Qwen3.5’s 35B doesn’t run on a phone the normal way, GLM-5.3’s cloud route isn’t local and may cost money, and the “10-minute” backup kit includes a 330 GB download.
Six free AI models and apps, the exact commands to run them on your own computer, and a backup kit for the ones you can keep on your own drive, where no company can switch them off. Everything except the cloud routes is local: once downloaded it runs with the Wi-Fi off, needs no account, and your chats stay on your machine. The exceptions are the giant GLM-5.3 and its Flash version, which a normal computer can’t hold, so their easy route runs on Ollama’s cloud, plus the optional gpt-oss:120b-cloud and the web search Claude’s app uses in section 6. Where the original guide got something wrong, it’s fixed and marked where it applies.
What held up, on 9 October 2026
Checked against Hugging Face’s records and the model cards and licences, ollama.com’s library and pricing, Ollama’s docs and source, Jan’s and AnythingLLM’s docs, Anthropic’s report and the news coverage of PewDiePie’s video.
What happened, and what didn’t
Around the start of September, Hugging Face disabled an unofficial copy of GLM-5.3 that a third party had uploaded with its safety refusals removed and branded for offensive cyber use; Z.ai’s own release stayed up. At the end of September, Anthropic published a report on how easily GLM-5.3’s safeguards come off. And PewDiePie says OpenAI suspended his account twice for “distillation” while he was building Ajax, his own AI, a fine-tune of Qwen3.5 9B. (The guide says a “stripped-down copy” was pulled in late September, and that OpenAI simply banned him.)
Neither story is quite “the cloud companies decide what you can use”, but the point underneath holds: a hosted service can suspend an account or remove a file under its own rules, and a model sitting on your own drive is the one thing nobody can switch off.
Everything below comes from its maker or from a major open-source project with thousands of GitHub stars; the models you pull with Ollama are its own repackaged builds of the makers’ files. The local models need no subscription and no API key, and your chats and files stay on your computer. The exceptions are the cloud routes, marked where they come up. (The guide calls everything here “the official version” and says your data never leaves your computer.)
Once a free model’s files are on your drive, they are yours to keep. Qwen3.5 and gpt-oss (Apache 2.0) and GLM-5.3 (Z.ai’s MIT-style licence) all allow commercial use, client work included, and they run offline. The cloud versions are a service, not a download.
Check what your computer can run
The number that matters is memory. On a Mac, RAM is shared with the graphics chip, so it is the only number: Apple menu, About This Mac, the Memory line. On a Windows PC, check Installed RAM in Settings, System, About, and your graphics card’s own memory too, since that decides speed. A model bigger than your graphics memory still runs, just slower. (The guide says RAM is the only number.)
Install Ollama, the engine for the commands below
The guide calls it the engine for everything else. Jan and AnythingLLM can also run models on their own; Ollama is what the commands on this page use.
Go to ollama.com/download
On a Mac or Windows, choose Download manually and install it like any normal app (Mac needs macOS 14 Sonoma or newer, Windows needs Windows 10 22H2 or newer). On Linux, paste the one-line install command shown on the page, also below. The page now leads with those one-line commands for every system; they do the same job.
Open your terminal
Mac: press Cmd + Space, type Terminal. Windows: press Start, type PowerShell.
Paste one command and press Enter
It downloads the model once, about 7 GB for this one, then you can chat with it right there. On an 8 GB computer, use qwen3.5:4b instead. To download without opening a chat, use ollama pull instead of ollama run.
Linux only: the install command from Ollama’s download page. Not in the guide.
Your first model
Type /bye to exit the chat. The model stays on your computer, so next time it starts without downloading again, even with the Wi-Fi off. (The guide says it starts instantly; after five idle minutes Ollama unloads it, and reloading from disk takes a few seconds.)
1. GLM-5.3 (Z.ai): the model behind the headlines
The model in the takedown story. (The guide calls it “the one they just pulled”; it wasn’t, only an unofficial modified copy was.) The official version is still free under Z.ai’s own MIT-style licence, business use included. It is very large, 744B parameters with 40B active, so here is the easy way and the owner’s way.
Easy way, any computer: sign in to Ollama (a free account at ollama.com), then run the cloud version. It is cloud-only on Ollama, with no local download.
First sign in. This opens ollama.com in your browser and returns straight away; finish signing in there before you run the next command. Not in the guide.
Then run the cloud model. The free plan covers a small monthly allowance on Ollama’s starter models, and GLM-5.3 may need pay-as-you-go credits ($1.40 per million input tokens, $4.40 per million output). This route sends your prompts to Ollama’s servers in the US and EU. Ollama says it doesn’t store them or train on them, but keep confidential client files on a local model.
The smaller GLM-5.3-Flash is cloud-only on Ollama too, and much cheaper ($0.15 per million input tokens, $0.50 output). Not in the guide.
Own the weights: the official files live at huggingface.co/zai-org/GLM-5.3, plus the smaller GLM-5.3-Flash. Download them to an external drive as your backup copy, but know the sizes first: GLM-5.3 is about 756 GB, and GLM-5.3-Flash (320B parameters, 18B active, plain MIT licence) about 330 GB. Neither runs on an ordinary computer; even Flash’s smallest community version needs a machine with 128 GB or more of memory. It is an archive, for the day you have the hardware.
Best for: coding, building websites and apps, long multi-step tasks. Z.ai calls it the most capable open-weights model for coding. The cloud version takes text only.
The base PewDiePie built on
The most flexible family, Apache 2.0, with sizes on Ollama from 0.8B (about 1 GB) to 122B (81 GB). PewDiePie’s Ajax is a fine-tune of this exact 9B. Swap 9b for 4b on an 8 GB laptop (about 3.5 GB), or for 27b (about 17 GB) or 35b (22 GB) on a strong machine. (The guide adds that “the 35B version even runs fully on a new iPhone”. It doesn’t, the normal way: the smallest sizes run on recent phones, and the 35B is a mixture-of-experts model, 35B in total with only 3B working at a time, so it is quick for its size on a computer with 32 GB.)
Best for: everyday writing, emails, summaries, reading images and screenshots. Every size accepts images; in the Ollama app you can drag one in.
OpenAI’s own free model
Apache 2.0 licensed, plus a one-line usage policy: follow the law. OpenAI says the 20B version runs within 16 GB of memory; it is a 14 GB download, so on a 16 GB machine close other apps first. It is strong at reasoning and code, with adjustable reasoning effort, but text only: use Qwen3.5 for screenshots. The bigger gpt-oss 120B needs about 80 GB of memory, or use gpt-oss:120b-cloud.
Best for: step-by-step reasoning, small apps and games, analysis.
ChatGPT with no internet
Download Jan from jan.ai
Free and open source (Apache 2.0), for Mac, Windows and Linux. No account needed. Jan has its own built-in engine and downloads its own copy of each model, so it doesn’t need Ollama, and it fetches a starter model on first launch. After that one-time download it works without internet.
Open the Hub
From the left sidebar. Pick a model that fits your memory: Jan labels each one Fits, May be slow or Won’t fit for your computer. Click Download. Prefer models from the maker’s own Hugging Face page or a well-known uploader.
Start a new chat
It looks and feels like ChatGPT. With a local model your chats stay on your machine and work offline. (The guide says everything stays on your machine; Jan’s optional cloud providers and its web search do go online.)
Best for: anyone who wants the ChatGPT experience without touching a terminal.
Chat with your own files
Download the desktop app from anythingllm.com/download
Free and open source (MIT) for Mac, Windows and Linux, no account needed. (The guide’s anythingllm.com/desktop link now redirects there.)
Choose Ollama as the LLM provider
Pick the model you already downloaded in Step 1, for example qwen3.5:9b. Ollama must be running; AnythingLLM finds it at http://127.0.0.1:11434.
Create a workspace and add documents
PDFs, contracts, proposals, SOPs, client briefs. Upload them to the workspace itself, so every chat can use them; files dropped into one chat only count for that chat.
Ask questions
It answers from your documents and shows a Sources list under each answer, naming the file it came from. It reads only the passages most relevant to each question, so for questions across every file, like the first two below, pin the key documents (the pushpin icon puts a whole document in front of the model) or ask about one document at a time.
Questions to try on day one
Best for: client docs and private business files you would never paste into ChatGPT. AnythingLLM sends anonymous usage counts by default (never the content); turn that off under Privacy in its settings.
Claude’s app, running your free models
Since Ollama 0.33 (August 2026), Claude’s desktop app can use Ollama models, with Ollama as its model provider: Claude keeps its interface, and the answers come from Ollama models on your Mac or on Ollama’s cloud. It isn’t a plugin or an MCP server; Ollama runs a small local gateway and points Claude’s app at it. Right now this works on Mac; Ollama says Windows support is coming soon. No Claude subscription is needed in this mode, and chats you have while connected are saved on your Mac rather than in your Claude history.
Download a model first
For example: ollama run qwen3.5:9b (or ollama pull qwen3.5:9b, to download without chatting).
Open the Ollama app and select Apps
This only works in the Mac app.
Connect Claude
Follow the prompts and restart Claude when asked. If Claude isn’t installed, Ollama offers to download it.
Switch models any time
Ollama Settings, Apps, choose your models, Restart Claude. You pick which Ollama model fills each of Claude’s model slots; one model can fill several. Restarting stops any task Claude is running.
To go back to normal Claude (your account and Anthropic’s models), disconnect it in Ollama Apps, or run this, then restart Claude. Quitting Ollama also switches Claude back, so keep Ollama open while you use it. Claude’s web search and auto mode go through Ollama’s online services in this mode, so they aren’t fully offline.
Ten minutes of setup for the essentials
The archive copy, step by step
Not in the guide. Hugging Face’s download tool, hf, needs Python 3.10 or newer, and its installer won’t install Python for you. A Mac’s built-in python3 is too old (3.9), so first install Python from python.org (or run brew install python). On Ubuntu or Debian you may also need sudo apt install python3-venv. Then install hf with this on a Mac or Linux:
On Windows, first install Python 3.10 or newer from python.org (tick Add Python to PATH) or the Microsoft Store, then run this in PowerShell:
Then, in a new terminal, list what will download, and download it to your drive. Change the path to your drive: on Windows something like D:\GLM-5.3-Flash. These are the maker’s original files (FP8 safetensors), not the GGUF builds Ollama and Jan load, so keep them as an archive.
Ollama keeps its downloaded models in a folder you can copy to an external drive: ~/.ollama/models on a Mac, /usr/share/ollama/.ollama/models on Linux, and C:\Users\<you>\.ollama\models on Windows. To keep them on the external drive permanently, open the Ollama app’s Settings and point Model location at a folder on the drive (Mac and Windows). On Linux, run sudo systemctl edit ollama.service and, under [Service], add an Environment line that sets OLLAMA_MODELS to the folder; then give the service the folder with sudo chown -R ollama:ollama /path/to/folder. Then restart Ollama. Not in the guide.
Only download models from the official maker pages or the official model links on this page. (The Lab’s uncensored-model entry, linked above, is a deliberate exception: it uses a community-modified copy, so the same caution applies.) Unofficial re-uploads of modified models can hide malware: researchers have found around a hundred malicious models on Hugging Face, most hidden in Python pickle files (.bin, .pkl, .pt) and some in Keras files (.h5). Prefer .safetensors or .gguf files from the maker’s own page, and keep Ollama updated. Unofficial copies can also be taken down at any time, which makes them a poor backup. (The guide says “cracked” models, but open models need no cracking. It also says modified copies “keep getting deleted”; Hugging Face disabled one GLM-5.3 copy, and dozens of others are still up.)
To make sure nothing ever goes to Ollama’s cloud, switch its cloud features off. In the Ollama app (Mac or Windows): open Settings and turn Cloud off. On Linux: run sudo systemctl edit ollama, add the two lines [Service] and Environment="OLLAMA_NO_CLOUD=1", save, then run sudo systemctl restart ollama; its log then shows “Ollama cloud disabled: true”. Cloud models and web search stop working; everything local carries on. Not in the guide.
Built from a beginner’s guide to keeping free AI models, kept in full and in its order, with its community sign-up left out and its corrections marked where they apply. Checked on 9 October 2026 against Hugging Face’s records for every repository named, the model cards and licence files, ollama.com’s library, tags and pricing, Ollama’s docs and source code, Jan’s and AnythingLLM’s docs, Anthropic’s 29 September report, and the coverage of PewDiePie’s 1 October video. Sizes, prices and versions change; Ollama was at 0.40.2 on the day.