Survival LLM: A Free Offline AI Assistant That Answers From Trusted Manuals
A free, open source app bundles a local AI model with first aid, FEMA and repair guides so you can ask questions with no internet after setup.
Imagine the power is out, your phone has no signal, and you need a first aid answer. Survival LLM tries to cover that. It is a free, open source app (MIT licence) for Mac and Windows that puts an AI model and a reference library on your own computer. After a one time setup that needs internet, it works fully offline, and your questions never leave the machine. It cites the source for each answer, so you can read the original passage. It reached more than 500 points on Hacker News this week.
The library is the interesting part. The full version is a 9 GB download (19 GB once installed) with medical references, FEMA and CDC guides, survival manuals, how to content and, in the larger edition, Wikipedia and iFixit repair guides. A smaller edition without those two is 1.3 GB. You can also add your own PDFs and notes, which the assistant searches locally. Which AI model you get depends on your computer’s memory. A Mac with 32 to 96 GB runs Qwen3.8 27B (an 18 GB download), Windows machines with 32 GB or more get NVIDIA Nemotron 3.5 Lightning, and a laptop with 16 to 31 GB gets IBM Granite 4.2 8B, which is only 6 GB. Setup takes 20 minutes to about 2.5 hours, mostly downloading, and the app is not from an app store, so your computer will show a security warning on first launch.
What’s behind it: this combines two ideas worth knowing. A local model is an AI that runs on your own hardware instead of a company’s servers. Pairing it with a library is called retrieval, which means the assistant looks up passages in trusted documents and builds its answer from them instead of from memory alone. That makes it easier to check and less likely to make things up, although it does not remove the risk.
The author’s own numbers show the trade offs. On 22 blackout and first aid questions, graded blind with an earlier build, the larger models scored 21 of 22 and the small Granite model 14 of 22. Smaller models make more mistakes, so check the cited sources. Answering also uses power, around 100 watts on a MacBook Pro, which the author says is roughly 100 to 300 answers per full battery.
What this means for you: for curious beginners, this is a gentle way to see what a local AI feels like, with no account and no cost. If you want it for emergencies, set it up now while you have internet, test it, and keep a printed first aid guide too. It is a helpful backup, not a replacement for emergency services.
Sources
A Tougher AI Scoreboard Widens the Gap Between Open and Closed Models
Artificial Analysis rebuilt its Intelligence Index with mostly private tests, and the gap between the best open models and the leaders grew from 3 points to 12.
