Offline AI assistant: how Survival LLM pairs a local model with a 930,000-article emergency library, and how people run AI with no internet

An offline AI assistant keeps both the language model and the material it consults on your own computer, so it keeps answering when the internet, or the cloud service behind a chatbot, isn't there. That idea drew wide attention on October 27, 2026, when Survival LLM, a free app that pairs a local model with a downloaded emergency library, was widely discussed on Hacker News. It runs small models of the kind covered in our guide to choosing a model for your laptop, so memory decides what you can run. This explainer covers what Survival LLM bundles, how the established tools handle offline use, and what an offline model can't do.

This explainer is based on the Survival LLM product page, the official documentation and project pages for Kiwix, Ollama, llama.cpp and LM Studio, and Google's Gemma 4 model card, checked October 28, 2026, and, for Survival LLM, its public GitHub repository, checked October 29, 2026.

The short version

• What it is: an app for Apple Silicon Macs and Windows PCs that downloads a model and a reference library once, then, its page says, "needs none" (no internet).
• What's in the library: about 930,000 of Wikipedia's most-read articles plus medical, emergency, repair and how-to sources. The full set is a 9 GB download; a 1.3 GB version skips Wikipedia and iFixit.
• Which model: one of four, chosen by your computer's memory, from a 6 GB download for 16 GB machines to 80 GB for Macs with 128 GB or more.
• License: the page says the app is open source under the MIT license; each library source keeps its own license. The page names no maker and its source-code links go nowhere, but a public GitHub repository that lists the page as its homepage publishes the code and every release.
• Limits: small models make more mistakes, the knowledge is frozen at download time, and none of its answers are verified medical or safety guidance.

What is Survival LLM?

Survival LLM describes itself as "An AI that still works when the internet doesn't": "a free, open-source AI assistant and survival library that runs on your own computer and works with no internet." You download a small app (212 MB for Mac, 25 MB for Windows), and it then fetches the model and the library. Setup takes 20 minutes to about 2½ hours depending on the computer and connection, the page says. "Set it up while you have internet; after that it needs none."

The design is retrieval: the model searches the local library before or after answering. Each answer lists the articles it read, which open with the passage highlighted, and is labeled as coming from the model's memory, from articles it read, or from search results it didn't open. You can add your own PDFs, Word files, ebooks and notes, which are searched locally. There's no account, and the page says "Nothing you ask leaves" your computer.

The project is very new. Its GitHub repository was created on October 24, 2026, the day version 0.1.0 was published there. Version 0.3.0, dated October 26, 2026, added a memory feature, and versions 0.3.1 through 0.3.4 all shipped on October 27. They added a "show reasoning" view and stopped some answers from opening with a made-up "Sources:" line. A refreshed library pack followed on October 28, which its release notes say removed about 2,700 repeated pages.

What does Survival LLM bundle?

Two things: a model matched to your hardware and a curated library.

Last verified: October 29, 2026

The page assigns IBM Granite 4.2 8B (6 GB) to machines with 16–31 GB of memory, Qwen3.8 27B (18 GB) to Macs with 32–96 GB, NVIDIA Nemotron 3.5 Lightning (25 GB) to Windows PCs with 32 GB or more, and NVIDIA Nemotron 3.5 Super (80 GB) to Macs with 128 GB or more. The repository's README also lists Gemma 4 E4B, a 10 GB download, as an optional faster choice for 16 GB computers. The four models' test scores, in that order 14 of 22, 21 of 22, 19 of 22 and 21 of 22, are maker-reported. The page says they count answers that got "the main point of official guidance right" on 22 blackout and first-aid questions, graded blind, on "an earlier build of the library." It doesn't publish the questions or say who graded them, and 22 questions is a small sample. The page doesn't state which quantization each model uses, and it names no license for the models themselves; for what "open" means for model weights, see our open-weight vs open-source explainer.

The library comes in two sizes: 9 GB to download and 19 GB installed, or 1.3 GB and 3 GB without Wikipedia and iFixit. Besides the Wikipedia articles, it lists WikiMed, MedlinePlus, the UK's Medicines A to Z, WikEM, Ready.gov (FEMA), the CDC's travelers' health pages, U.S. military medicine and survival manuals, wikiHow, iFixit, Appropedia and Stack Exchange sites on home repair, cooking, gardening, the outdoors and ham radio. The page calls these "openly licensed (some for non-commercial use only) and public-domain" and shows each source's license with its articles.

The product page doesn't name the engine or the file formats; it says only that the Mac app includes "the AI engine" and the Windows app downloads it. The repository's README is more specific: the app runs its models through Ollama, bundled with the Mac app and downloaded on first use on Windows, and builds its library from Kiwix ZIM files, converted into plain-text files with a SQLite full-text search index.

What hardware does an offline AI assistant need?

Mostly memory. Survival LLM needs an Apple Silicon Mac (M1 or newer) on macOS 14 Sonoma or later, or 64-bit Windows 10 or 11, with 32 GB of RAM for the Windows-only Nemotron model. Disk space is the model's size plus 5 GB, plus about 27 GB during a full library install (10 GB for the small library).

The tiers follow the arithmetic in our laptop model guide: a model's weights take roughly parameters × bits per weight ÷ 8 bytes, and the conversation's context adds more on top. As an illustrative estimate using that rule, a 27-billion-parameter model at about 4.5 bits per weight needs around 15 GB for its weights, and an 8-billion-parameter one around 4.5 GB. Survival LLM's listed downloads, 18 GB and 6 GB, are a little larger, as real 4-bit builds usually are. The rest of the machine's memory goes to the operating system, the app and the context.

Speed and power matter offline too. The page says answers take 5 to 40 seconds, that a MacBook Pro draws about 100 W while answering, and that a full battery gives "roughly 100 to 300" answers. A dedicated neural processing unit doesn't change much here, as our NPU explainer notes: the common local runtimes use the CPU and GPU. Small boxes built to run big models locally, such as Lenovo's AI Host P7, are aimed at the same need, but its specs aren't confirmed by Lenovo.

How do people run AI with no internet?

Survival LLM packages a pattern that already exists in separate tools. Each needs a connection once, to download content or models, and the docs of each describe offline use after that.

Last verified: October 28, 2026

Kiwix is the reference library half. The Swiss nonprofit describes itself as "making free knowledge accessible where the Internet is not." Its reader opens ZIM files, "highly compressed archives of web content," on Windows, macOS, Linux, Android and iOS: "Install Kiwix once, download the content of your choice then access them anytime, without connection." Wikipedia comes in three versions: mini (introductions and infoboxes), nopic (full articles without images) and maxi.

Ollama is the most common way to run the model half, as our Ollama guide shows. Its FAQ says "Ollama runs locally. We don't see your prompts or data when you run locally," and that it "can run in local only mode by disabling Ollama's cloud features," by setting OLLAMA_NO_CLOUD=1 or disable_ollama_cloud in its server config. You give up its cloud models and web search. A model file copied from another machine can be imported with a Modelfile line, FROM /path/to/file.gguf, and ollama create.

llama.cpp, the MIT-licensed engine many local apps build on, aims "to enable LLM (and VLM) inference with minimal setup." Its server takes a model path directly, and its --offline option "forces use of cache, prevents network access."

LM Studio says that once a model is on your machine, "you should be good to go entirely offline," and that for document chat "nothing you upload into LM Studio leaves the application." It recommends 16 GB of RAM on Macs and Windows PCs, plus at least 4 GB of dedicated video memory on Windows.

What are the limits of offline AI?

Four stand out.

• It still makes things up. Survival LLM's page says plainly: "AI models make mistakes, the small ones more often." That's why it shows its sources. Model makers say the same: Google's Gemma 4 card warns that its models "may generate incorrect or outdated factual statements."
• Its knowledge is frozen. A model knows only what was in its training data. The Gemma 4 card gives a cutoff of January 2025. Offline, nothing fills the gap: no web search, no news, and the library is only as current as its last download.
• Nothing updates by itself. New model versions, library refreshes and app fixes all need a connection. Survival LLM shipped five versions in two days, and an offline copy stays on whichever one it had.
• It needs power and a capable computer. A model that fits only a 16 GB machine is the smallest and, on the maker's own test, the least accurate. Each answer draws real power, which matters most in exactly the situations the app is built for.

Readers on Hacker News raised similar doubts. Some questioned whether an LLM adds much over searching a plain offline library like Kiwix, and others pointed to hallucination.

Can you rely on AI for medical or survival information?

Not as verified guidance. Survival LLM's library draws on public health and emergency sources, but the page doesn't say that the model's answers are reviewed by clinicians or emergency agencies, and it carries no medical disclaimer beyond its warning that models make mistakes. The maker's own test found the smallest model right on 14 of 22 emergency questions. What the app does provide is a link from each answer to the source passage it used, so the original text is available alongside the AI's summary.

What's confirmed, reported and uncertain

• Confirmed (product page): the models and download sizes by tier, library contents and sizes, system requirements, power and timing figures, offline-after-setup design, version history and the stated MIT license.
• Per the GitHub repository (which lists the product page as its homepage, though the page doesn't link back): the public code, the MIT license text, the release history, the Ollama engine and the ZIM-based library build.
• Confirmed (project docs): how Kiwix, Ollama, llama.cpp and LM Studio behave offline.
• Maker-reported: the emergency-test scores.
• Uncertain: who makes Survival LLM (the product page names no maker), whether the GitHub repository is the official one (the page's source links still pointed nowhere on October 29), and the licenses of the bundled models.