Skip to content
Week 44News, and what it means for you

AI models

Survival LLM: a free AI assistant that works without the internet

An open-source app puts an AI model and an emergency library of 930,000 Wikipedia articles on your own computer. Once installed, it needs no internet at all.

How long does food in the fridge stay safe in a power cut? A free app answers questions like that without the internet, using an AI model and an emergency library that both live on your own computer.

Survival LLM appeared on GitHub on 24 October and took off on Hacker News on 27 October, collecting 539 points and more than two hundred comments. Its maker goes by Brangus on GitHub and says nothing else about themselves. The code is open source under the MIT licence, and ten versions came out within a week, adding among other things a memory for personal notes.

The library holds some 930,000 of Wikipedia’s most-read articles, medical references such as WikiMed and MedlinePlus, emergency guides from Ready.gov and the CDC, the US public health agency, military first-aid and survival manuals, and practical help from wikiHow, iFixit and Stack Exchange. Everything together is a download of about 9 GB, or 1.3 GB without Wikipedia and iFixit. You can add your own PDFs, Word files and notes, such as the manual for your generator.

Which model you get depends on your computer’s memory. A Mac with 128 GB or more gets NVIDIA Nemotron 3.5 Super (80 GB), a Mac with 32 to 96 GB Qwen3.8 27B (18 GB), a Windows PC with at least 32 GB Nemotron 3.5 Lightning (25 GB), and a computer with 16 to 31 GB IBM Granite 4.2 8B (6 GB). Every answer names the articles it came from, and you can open them with the passage used highlighted.

Reactions on Hacker News were mixed. Someone who had lived through a category 5 hurricane wrote that something like this would have been a blessing. Others asked why you wouldn’t just use a searchable archive or books, and where the power comes from when everything is down.

AI models make mistakes. Every answer shows its sources so you can check.

From the Survival LLM documentation on GitHub

Offline Wikipedia is nothing new

In 2006 Emmanuel Engelhart made a version of Wikipedia that works without the internet and called it Kiwix. Kiwix packs entire websites into a single file, a ZIM file, and is used in remote villages, refugee camps and on submarines. Survival LLM builds much of its library from such ZIM files, about 50 GB in all, and turns them into plain text with a search index.

How a big model fits on a laptop

A language model consists of billions of numbers. Qwen3.8 27B has 27 billion of them; at 16 bits per number that would be about 54 GB. By storing each number in roughly 4 bits, which is called quantisation, the model fits into 18 GB. macOS only lets the graphics side of a Mac use about three quarters of its memory, which is why the largest model, at 80 GB, needs a Mac with 128 GB.

Open models are improving fast. IBM’s chief executive, Arvind Krishna, told investors on IBM’s earnings call, according to The Register, that they are now about 90% as good as the best closed models, at 10 to 25% of the price.

Answers from the library

The model does not answer from memory; it searches first. The app cuts the library into passages of about 1,500 characters and searches them by keyword, with official sources ranked higher. The model reads what it finds and answers with the sources attached.

The maker tested eight models on 22 questions about power cuts and first aid. Other AI models graded the answers blind, against official guidance from the CDC, the NHS and the Red Cross, among others. Nemotron 3.5 Super got the main point right 21 times, Granite 4.2 8B 14 times. The maker openly adds that the test ran on an earlier version of the library.

Our analysis

The clever part is the library, not the model

What makes Survival LLM interesting is not that a model runs without the internet. It is how it is put together: the model is the librarian, not the encyclopaedia. It searches a fixed collection, names its sources, and the app checks those sources as well. That answers the question of why you wouldn’t just search: searching works when you know the right words, and a model helps precisely when you don’t. The critics are right, though, that the model can still get things wrong.

What most reports miss is the maker’s honesty. They publish the dangerous answers, the made-up sources and which models were dropped as a result. That is the lesson for anyone putting AI to work: test with real questions and build in checks. And don’t mistake a good test score for a good decision. In a 2025 study with 1,298 participants, models on their own identified the right condition in 94.9% of cases, but people working with a model did so in fewer than 34.5%, no better than the group that chose its own source.

For organisations, this shows the principle is coming of age: AI that runs on hardware you control, keeps working without the internet and answers from your own documents, with the source attached. Our expectation: for manuals, procedures and case files this will become the normal choice, especially where data must not leave the building.

What stands out

  • The dangerous answers are published too Granite 4.2 3B advised against putting a tourniquet on an arm and said to stay 3 feet, under a metre, from a downed power line, while the guidance is 35 feet, over 10 metres. The maker does not offer that model for this reason.
  • Models make up sources In a test of 14 questions, the largest model six times named a source it had never opened. The app now replaces such a sources line with the articles the model actually read.
  • Bigger is not always better Qwen3.8 27B, at 18 GB, got the main point right as often as the 80 GB model: 21 times out of 22. An answer did take 40 seconds instead of 12.
  • Power is still the limit On a MacBook Pro the app draws about 100 watts while answering. A full battery is good for roughly 100 to 300 questions, according to the maker.

What it means for you

AI that works without the internet and answers from your own documents is no longer an experiment. Whether you can rely on it depends on the model, the documents and how well you test.

  • Keep sensitive questions in-house When the model runs on your own hardware, questions and documents never leave the building, and it keeps working when the internet falters.
  • Have it answer from your own documents Your colleagues can check a model that searches your manuals and procedures and names the source. They cannot check one that talks from memory.
  • Test with your own questions Draw up a list of real questions whose answers you know, as the maker did with 22, and pay most attention to answers that are dangerously wrong.

Where we come in · Solutions

We put AI in your own office

We supply AI servers for the office: an Apple Mac Studio, cut off from the internet, running open models such as Mistral, Gemma and gpt-oss, so sensitive data never leaves the building. And with our AI knowledge base, your colleagues ask questions about your manuals and instructions in plain language and get an answer straight away, with the source attached.

Which questions would you rather keep in-house?

Tell us which documents you work with. We’ll help you think through AI on a server of your own and always reply within one working day.

More from week 44

All news

Search our services, the news and the rest of the site.