Survival LLM puts an emergency library on your laptop

Survival LLM wants to make an emergency reference library searchable through a chatbot on your own computer, even when the internet is unavailable. Its October 28 v0.3.4 release adds explicit reasoning budgets and fixes search behavior, giving an existing community project a concrete new update as its offline AI pitch draws debate on Hacker News.

This is a standalone desktop app, not a newly trained model. The project supports Apple Silicon Macs on macOS 14 or later and 64-bit Windows 10 or 11, using Ollama for local inference. Users first download the application, engine, model and library while online. The author says it then works offline without a cloud API account. Models and reference packs require substantial additional downloads and storage; the small application download is not the complete installation.

The new release replaces the earlier thinking control with selectable reasoning budgets. Its notes also describe retrying with the best articles already found when a search produces no answer, honoring explicit requests to search, and distinguishing information from model memory, opened articles and unopened search results. Those are documented project claims, not independently measured reliability results.

The library can also index a user’s text documents, including PDFs, Word files and EPUBs; scanned pages and photos are not supported by that workflow. The code is MIT licensed, but the bundled reference material is not uniformly permissive: the source list includes materials with noncommercial restrictions. Commercial reuse requires checking the underlying licenses.

The community debate is part of the story. English Hacker News comments argue both for using models to locate relevant reference passages and against trusting generated emergency instructions. The README itself documents hallucinated citations and limitations in the author’s evaluations, including differences between an older test library and the public pack. Offline access is a practical capability; it does not establish that an answer is safe medical advice or that the system will be reliable in an emergency.