Softwareaiofflinelocal-llmopen-sourcepreparedness
Survival LLM Puts 930,000 Wikipedia Articles, a Medical Encyclopedia and Military Survival Manuals in a Free 9 GB Offline AI That Picks Its Model by Your RAM
The free, open-source app runs a local AI model over offline Wikipedia, WikiMed, U.S. military medicine and survival manuals and iFixit guides, and its makers measured about 100 W per answer on a MacBook Pro, roughly 100 to 300 questions per battery.
The pitch is a library first and a chatbot second. “Survival LLM is a free AI assistant and emergency library that runs entirely on your own computer,” the page says. The AI answers from what’s on your disk: the Wikipedia subset, WikiMed, the military manuals, the repair guides. Everything is “a 9 GB download for everything (19 GB once installed),” and a 1.3 GB version leaves out Wikipedia and iFixit for small drives.
The model choice is automatic. “It checks your computer’s memory and downloads the AI model that fits. More memory means a smarter model.” A Mac or PC with 16 to 31 GB gets IBM Granite 4.2 8B, a 6 GB download. A Mac with 32 to 96 GB gets Qwen3.8 27B; a Windows PC with 32 GB or more gets NVIDIA Nemotron 3.5 Lightning; a Mac with 128 GB or more gets NVIDIA Nemotron 3.5 Super, an 80 GB download.
The page publishes its own test scores, and they make the case for more RAM better than any marketing line could. On 22 blackout and first-aid questions, graded blind on whether the answer got the main point of official guidance right, the two big Mac models scored “21 of 22 right.” Nemotron 3.5 Lightning scored 19. The 8B model on a 16 GB machine scored 14. The page’s own warning: “AI models make mistakes, the small ones more often,” and the scores come from “tests on an earlier build of the library.”
Then there’s the power bill. “We measured about 100 W while it answers,” the page says, at 5 to 40 seconds a question. On a MacBook Pro, its makers estimate a full battery “answers roughly 100 to 300” questions. The other catch is timing: “Set it up while you have internet.” The page names no person behind it, and its GitHub link is a placeholder for now, so the code the MIT license covers isn’t one click away yet.
Expect more of these: local models wrapped around a fixed, checkable library, sold on what they do without a connection. The open question is who keeps the library current, since a survival manual from 2026 and a Wikipedia snapshot age the moment the download finishes.