Survival LLM is a desktop application that packages a locally-running large language model with a curated emergency reference library — medical guides, FEMA protocols, CDC guidance, military survival manuals, iFixit repair guides, and roughly 930,000 Wikipedia articles. Once installed, it requires zero internet connectivity. Every answer cites its source, with the relevant passage highlighted for manual verification. The tool auto-selects a model sized to your hardware. At the top end, a 128 GB Mac runs NVIDIA Nemotron 3.5 Super (80 GB download, 21/22 on the project's emergency test), while a 16 GB machine gets IBM Granite 4.2 8B (6 GB download, 14/22). The scoring methodology is transparent if rudimentary: 22 blackout and first-aid questions, graded blind. Smaller models make more mistakes, and the app says so plainly. The library itself is a 9 GB download (19 GB installed) spanning MedlinePlus, WikiMed, WikEM emergency medicine, Ready.gov, UK Medicines A to Z, Appropedia, wikiHow, Stack Exchange Q&A on home repair, cooking, gardening, ham radio, and the outdoors. A slimmer 1.3 GB version drops Wikipedia and iFixit. Users can add their own PDFs, Word documents, and ebooks — a generator manual, a family emergency plan — and the AI searches those too. Privacy is structural, not policy. There is no account creation, no telemetry, no data leaving the machine. The entire system is free and open source, hosted on GitHub. This is not a freemium funnel — there is no premium tier, no API call, no cloud dependency. The value proposition is that the tool works precisely when commercial alternatives do not. Power consumption is manageable for emergency scenarios: roughly 100 watts while answering, sustained for 5 to 40 seconds per question. A full MacBook Pro battery yields an estimated 100 to 300 answers. A small generator or portable power station extends that indefinitely. The latest update (v0.3.4, October 2026) adds configurable reasoning depth, memory notes for persistent context like family allergies, and cleaner source attribution. The project sits in a small but growing class of resilience-oriented software — tools built for degraded conditions rather than optimized for peak connectivity. It does not compete with cloud AI on capability; it competes on availability. The tradeoff is explicit: you get a dumber model that actually works when nothing else does. The honest limitation is model quality at the low end. At 14/22 on emergency questions, the 8B model is wrong roughly a third of the time on critical health and safety topics. The mitigation — always showing sources — is real but requires a user willing to read and verify, which is not guaranteed in a crisis. The tool is most valuable to the kind of person who would have kept a shelf of reference books anyway.
