Survival LLM: offline AI with an emergency library and cited sources

What it's about

Survival LLM is a free, open-source desktop app (MIT license) that combines an AI assistant and a survival library on your computer. After setup, it works without internet. You can ask, for example, how to keep food safe during a power outage, how to treat a burn or how to purify water. The app searches a local library, answers from what it finds and shows the passage the answer comes from. There are versions for Macs with Apple Silicon (macOS 14 or newer) and for Windows 10/11 (64-bit). A Portuguese report names version 0.3.4 from October 27, 2026 as the current release.

What the app can do

• Offline and private: The model and library are stored locally. The server only listens on 127.0.0.1.
• Sources: Every answer lists the articles that were read. One click opens the article with the passage highlighted, plus its license.
• See the reasoning: "Show reasoning" reveals what the model thought, searched for and read before answering. A thinking budget (off, 1,000, 4,000, 16,000 tokens or unlimited) controls the effort. Tokens are roughly the word fragments a model processes.
• Customizable: The instructions given to the model are editable, as is a length limit for answers.
• Your own files: PDFs, Word documents, EPUBs, web pages and text files (such as a generator manual or an emergency plan) are searched too. The app can't read scanned pages or photos yet.
• Memory: If you say your daughter is allergic to penicillin, the app saves a note. You can undo or edit it. Saving can be turned off.
• Updates: Once a day, the app checks a small file on the project site without sending any data about you, and updates itself if you want.

The library

It includes about 930,000 of the most-read Wikipedia articles, WikiMed, MedlinePlus, a British drug guide, WikEM, Ready.gov, CDC, US military and FEMA manuals, the USDA canning guide, risk sources for micromort estimates (life tables, US death rates), wikiHow, iFixit, Appropedia, Energypedia and practical Q&A pages. For search, the app uses SQLite full-text search (FTS5, BM25 ranking) over passages of about 1,500 characters, with official sources weighted higher. The models have three tools: search_library, read_article and show_article. Each source keeps its own license, and some are for non-commercial use only.

Which model you get

The app checks your RAM and suggests a model. It uses Ollama as its engine: built in on the Mac, downloaded with one click on Windows (1.6 GB).

Gemma 4 E4B (10 GB) is offered as a faster option for 16 GB machines. With 8 GB, you can only search the library, because no tested model was small and safe enough.

Model test results

Each model answered 22 questions about blackouts and first aid. Independent AI judges scored the answers blind against official guidance from USDA, CDC, FEMA, EPA, NHS, WHO and the Red Cross. The tests ran on an earlier library version, on an M4 Max MacBook Pro with 128 GB.

Granite 4.2 3B isn't offered: it advised against a tourniquet on an arm and gave 3 feet as the safe distance from a downed power line (the guideline is 35). Qwen 3.6 is missing too, because it looped on two questions, once for 94 minutes. That's why there is a length limit.

Made-up sources

Models sometimes cite sources they never saw. In a test with 14 questions, Nemotron Super did this six times, Lightning three times, Gemma three times, Granite twice and Qwen3.8 27B once. So the app compares the "Sources" line with the articles actually opened and replaces false entries with the real list.

What you need to try it

• A Mac with Apple Silicon (macOS 14+) or Windows 10/11 64-bit, with at least 16 GB of memory for a model.
• Storage: model size plus 5 GB, plus about 27 GB temporarily for the full library (21 GB afterward) or around 10 GB for Essentials (3 GB afterward).
• Internet only for setup. It takes 20 minutes to about 2.5 hours, with about 5 minutes of actual clicking.
• The app isn't signed by Apple or Microsoft. On a Mac, you have to allow it under Privacy & Security with "Open Anyway"; on Windows, click "More info" and then "Run anyway".
• According to the report, a MacBook Pro draws about 100 W and takes 5–40 seconds per answer. A full battery charge should last for roughly 100–300 questions.

Our take

The app solves a real problem: emergency knowledge usually lives online. Showing sources makes sense, because small models get things wrong noticeably more often. The project itself says AI makes mistakes, and the numbers back that up: the 6 GB model missed the key point on 8 of 22 questions. The tests ran on an earlier library version and include material that isn't in the public downloads, so they're a model comparison, not a guarantee. For medical or safety-critical questions, read the source yourself. It is no substitute for emergency services or professionals.