An AI assistant that keeps working when the network doesn't. It is Survival LLM, a free, open-source app that runs entirely on your computer and answers emergency questions: how to keep food safe during a blackout, how to treat a burn, how to purify water. Every answer lists the articles it drew on, and clicking one opens the passage, highlighted.
The setup is simple. The AI model and the reference library live on the computer, so nothing you ask leaves the machine. Installation needs a connection; afterwards the program doesn't. Setup takes from about twenty minutes to two and a half hours, depending on the computer and the network speed, but the user only clicks for a few minutes.
The library holds plain-text copies of openly licensed and public-domain references, searchable in a fraction of a second: about 930,000 of the most-read Wikipedia articles, WikiMed, MedlinePlus, the FEMA emergency guides on Ready.gov, the CDC travelers' health pages, U.S. military medicine and survival manuals, wikiHow, the iFixit repair guides and Stack Exchange sections on cooking, gardening and ham radio.
The model depends on the machine. On first launch the app checks the computer's memory and downloads the AI that fits. According to the developers' figures, a Mac with 128 GB or more uses NVIDIA Nemotron 3.5 Super, which got 21 of 22 right in a blind emergency test, while a Windows PC with at least 32 GB uses Nemotron 3.5 Lightning, with 19 of 22. Below that, smaller models make more mistakes: IBM Granite 4.2 8B got 14 right. That is why the app always shows its source to check.
The app is 212 MB on Mac and 25 MB on Windows; the full library takes 19 GB once installed (1.3 GB, or 3 GB installed, without Wikipedia and iFixit). It needs an Apple Silicon Mac on macOS 14 Sonoma or later, or 64-bit Windows 10 or 11. It declares no account and no tracking, and users can add their own PDFs, manuals and family plans, which stay on the computer.
Power use is another practical point. On a MacBook Pro the developers measured about 100 W while answering, for 5 to 40 seconds per question: a full battery covers roughly 100 to 300 answers, and a small generator can run it easily.
Version 0.3.4, released on October 27, 2026, adds a setting that limits how much the AI may reason before answering, a memory feature for personal details such as an allergy, and the option to see the model's reasoning under each answer. The project is hosted on GitHub, where an independent fork maintains its own updates.
No comments yet. Be the first!