Survival LLM: AI assistant and emergency library that runs on your own computer
- ID
- 40319
- Status
- summarized
- Published
- 28 Oct 2026, 3:55 AM
- Fetched
- 28 Oct 2026, 6:36 PM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://www.brangus.io/survival-llm
- Source URL
- https://hnrss.org/best
Summary
- Score
- 7.5
- Created
- 28 Oct 2026, 6:37 PM
- Tags
- Audience
- developersvibe_codersai_ml_learners
What happened
Survival LLM is a free, open-source offline AI assistant and emergency library for macOS and Windows that answers first-aid and blackout questions from local sources and highlights the exact passages used. It ships a 9 GB library (19 GB installed) with about 930,000 Wikipedia articles plus FEMA, CDC, medical, and repair references, or a 1.3 GB version without Wikipedia and iFixit, and selects an on-device model by RAM from IBM Granite 4.2 8B at 6 GB up to NVIDIA Nemotron 3.5 Super at 80 GB. In the project's own 22-question test, the smallest model got 14/22 right versus 21/22 for the larger models, and setup takes 20 minutes to about 2.5 hours.
Why it matters
If you build for users with unreliable connectivity or privacy constraints, this is a concrete reference for offline-first RAG: bundle the corpus, pick the model by available RAM, and surface citations instead of a chat-only answer. The 14/22 versus 21/22 result is a warning that the smallest supported model may be too weak for safety-critical guidance, so decide deliberately whether to gate features by hardware or show stronger warnings.
Discussion angle
The project's own test has the 8B model at 14/22 on first-aid questions—would you ship an offline assistant with that accuracy if it always cites sources, or is that a hard no for safety-critical use?