The situation it's built for
In a blackout with no internet, questions like "how long is the food in the fridge safe", "how do I treat a burn" or "how do I purify water" are exactly when online AI assistants stop working. Survival LLM's approach is to install the model and a reference library on your computer while you still have a connection, then run fully offline: no account, no tracking, and questions never leave the machine. Developer anonnerdfrenzy created the GitHub repository on October 24 and shipped five releases, 0.3.0 through 0.3.4, on October 26 and 27. The project page calls it free and open source, but the license on GitHub isn't marked as a standard open-source license, and each source in the library keeps its own license, some for non-commercial use only.
A model sized to your memory, with sources on every answer
On first launch the app checks your computer's memory and downloads the matching model:
- Macs with 128 GB or more: NVIDIA Nemotron 3.5 Super, 80 GB;
- Macs with 32–96 GB: Qwen3.8 27B, 18 GB;
- Windows PCs with 32 GB or more: NVIDIA Nemotron 3.5 Lightning, 25 GB;
- Macs or PCs with 16–31 GB: IBM Granite 4.2 8B, 6 GB.
The library is searchable plain text, including about 930,000 of Wikipedia's most-read articles, WikiMed, MedlinePlus, the UK Medicines A to Z, WikEM emergency medicine, FEMA's Ready.gov guides, CDC travelers' health, US military medicine and survival manuals, wikiHow, iFixit repair guides and Appropedia. The full download is 9 GB (19 GB installed); a smaller version without Wikipedia and iFixit is 1.3 GB. You can also add your own PDFs and Word documents, such as a generator manual or a family emergency plan.
The author blind-graded 22 blackout and first-aid questions on whether each answer got the main point of official guidance right. The top two model tiers each got 21 right, Nemotron 3.5 Lightning got 19, and Granite 4.2 8B only 14. That's why the app lists the articles it read under every answer, and clicking one opens the source with the cited passage highlighted. On a MacBook Pro the author measured about 100 W while answering, 5 to 40 seconds per question, so a full battery handles roughly 100 to 300 questions.
How to read it
This is a personal project, the test has only 22 questions, and it was run on an earlier build of the library, so it shouldn't be treated as a medical advice tool. Its real value is packaging "local model plus traceable retrieval" into an app ordinary people can install. Offline answers may not be perfect, but being able to open the FEMA or MedlinePlus original directly is more reliable than taking the model's paraphrase on trust. If you want it, set it up while you're online and try a few questions you already know the answers to, to get a feel for how well the smaller models do.
via: Survival LLM project page, GitHub: anonnerdfrenzy/survival-llm