Survival LLM: The Offline Emergency AI With a 930,000-Article Library, Tested on 22 Questions

Survival LLM is a free app that puts a small language model and a 930,000-article emergency library on your own computer, so you can ask "how long is fridge food safe in a blackout?" with the internet off. It reached the Hacker News front page on October 27, 2026, with a discussion of about 49 points. The project page and GitHub repository are unusually candid about how often the models get things wrong, which makes it a useful case study in where local AI is and is not ready for high-stakes questions.

The short version: the app is real, free and well documented, and its own test says the big models are decent and the small ones are not. The commenters' central objection, that a searchable archive would do the job without an LLM, deserves a serious answer.

TL;DR: what readers ask first

What does Survival LLM actually do?

You ask a question; a local model, run through Ollama, calls three tools. It searches the library, reads an article, and shows the article on screen with the cited passage highlighted. Search is a SQLite full-text index (FTS5, BM25 ranking over passages of about 1,500 characters), with official sources ranked higher. The server runs on 127.0.0.1 only.

The library is the real asset. The page lists about 930,000 of Wikipedia's most-read articles, WikiMed, MedlinePlus, WikEM, Ready.gov, CDC travelers' health, U.S. military medicine and survival manuals, wikiHow, iFixit repair guides, Appropedia and Stack Exchange Q&A on home repair, cooking, gardening and ham radio. The repository counts about 1.5 million articles in total. You can also add your own PDFs, Word files and ebooks, such as a generator manual or a family emergency plan, and the AI searches them too. The app can also save short memory notes ("allergic to penicillin"), stored locally with an undo.

Which model does your computer get?

The app checks memory and downloads the model that fits. These are the author's numbers from a blind test.

Models the author tested but does not offer include Granite 4.2 3B, which got 8 of 22 right and had 5 harmful answers, two of which "could kill." One example in the README: it told people not to use a tourniquet on an arm and to stay 3 feet from a downed power line, where the guidance is 35. Qwen 3.6 was dropped after getting stuck repeating itself on two questions, once for 94 minutes. Computers with 8 GB can search the library but get no recommended model.

If you run local models already, this overlaps with our coverage of Nemotron 3.5 Lightning fine-tuning and the offline coding agent with OpenCode, Ollama and Qwen3.5 9B.

How was it tested, and how far should you trust it?

Each model answered 22 blackout and first-aid questions covering food, generators, water, carbon monoxide, hypothermia, bleeding, burns, insulin, downed lines, fallout and botulism, using the full app with its prompt, tools and library search. AI judges graded every answer blind, against USDA, CDC, FEMA, EPA, NHS, WHO and Red Cross guidance: one for accuracy, one for safety, and a skeptic re-checking anything flagged harmful.

Read the caveats the author states himself. The tests ran on an earlier library build, before Wikipedia was added, and included some material not in the public downloads. The prompt has gained rules since, including one for carbon monoxide. The author says to "treat the table as a comparison between models, not a guarantee for the version you download." The judges are AI, graded against human guidance, and the sample is 22 questions, so differences of one or two answers are not meaningful.

Two findings are worth quoting as lessons. First, some models kept searching until they ran out of lookups and never answered; the app now tells them to answer from what they read. Second, several models gave "if you must" exceptions for running cars and generators near the house; the prompt now rules those out.

Made-up sources: the weak point

The most instructive section of the README is about citations. Each answer ends with the model's own "Sources:" line, but models name sources they never opened. In a separate 14-question check dated October 24, 2026, answers named a source the model had neither opened nor seen in its results: Nemotron Super 6 times, Lightning 3, Gemma 3, Granite 2 and Qwen3.8 27B once. The app now checks that line against what the model actually opened and replaces any line that names more with the list it really read.

This matters beyond survival. A "cited" answer is not automatically a grounded one, and the fix here, comparing claimed citations to retrieval logs, is a pattern any retrieval app can copy. It echoes the issues we covered in AI text that is hard to read and trust.

What developers are saying

The Hacker News thread was supportive of the idea and sharply split on the design.

• Why an LLM at all? User rnd0 asked why not gather the resources into an archive searchable without an LLM, citing hallucination and hardware needs, and pointed to Kiwix. TheRealPomax made the same point: a classic search bar over the same data would be smaller and as trustworthy as its sources.
• The accuracy table. saaaaaam noted that the 16-31 GB tier is wrong nearly half the time and suggested simply buying books instead. walrus01 argued that a hallucination-prone model is a questionable way to distribute survival information, and suggested streaming a much larger model from disk instead of fitting one in memory.
• Hardware demands. anonbeat wanted it smaller, and TGower was disappointed that Gemma E4B scored poorly, hoping a cheap phone could one day carry this. doublerabbit asked how it would run without electricity; the project page says a small generator or power station runs it, measuring about 100 W per answer on a MacBook Pro.
• Trust. abujazar pointed to the unsigned-app warning and distrusted the app; the page explains the warning arises because it is not sold through an app store.
• Alternatives. yoble found Project Nomad, and mentalgear suggested a plain vector search that runs on any computer. armcat described a 2018 offline project using Wikipedia, an inverted index and TextRank on a Raspberry Pi.

These are commenters' views, not verified findings, but the pattern is consistent: the library is the valuable part, and the model is the contested part.

Should you install it?

Install it if you already want a local, private model with retrieval over your own documents, and you have a Mac with 32 GB or more or a Windows PC with 32 GB or more. On those machines the author's numbers are respectable, and the cited-passage view lets you verify each answer.

Be cautious on 16 GB machines. A model that gets 14 of 22 main points right is a reading aid, not a substitute for official guidance, and the app's own README says so. The practical rule: use it to find and open the source article, then read the source yourself. For life-threatening situations, a printed first-aid manual and local emergency services remain the baseline.

Setup takes 20 minutes to about 2.5 hours depending on your connection, with about 5 minutes of clicking. On first launch macOS and Windows warn that the app is not from a store; the page gives the steps to approve it. If that makes you uneasy, run from source with the Python instructions in the repository, which uses only the standard library plus Ollama, pywebview, pypdf and truststore.

Why it matters

Three takeaways stand out. Local retrieval plus a small model can be packaged for non-experts. Honest evaluation, with harmful-answer counts and named failures, is rare and worth copying. And the commenters' challenge, that retrieval alone may beat generation for safety-critical lookup, is the right design question for any offline assistant, including the ones in our coverage of Opera's proposed WebAI standard for local models and Liquid AI's LFM3 family.

Related reading

• Offline coding agent with OpenCode, Ollama and Qwen3.5 9B
• Nemotron 3.5 Lightning local fine-tuning guide
• Opera WebAI: local models as a web standard
• Liquid AI LFM3 open-weight family
• Six patterns that make AI text hard to read

Details are accurate as of October 27, 2026, based on the project page, README and Hacker News thread; test numbers are the author's own.