For teams assessing open-source assistants, Survival LLM is worth reviewing as a local-first emergency knowledge base: the Mac download is 212 MB, the Windows download is 25 MB, and the source code is free and open source. It answers blackout, burn, and water-purification questions from bundled references after a one-time setup, and it shows the source material with the relevant passage highlighted. The engineering signal is not chat quality alone; it is whether model selection, disk footprint, offline operation, and source visibility can be verified on a real machine.

The model and library are stored on the user's computer, so queries do not need an internet connection after setup and nothing asked by the user leaves it.
Each reply can show the articles it read; opening one lets the user read the article and see the exact highlighted passage.
The bundled collection is aimed at emergencies and practical repair: medical references, FEMA and CDC guidance, survival manuals, how-to guides, repair guides, and about 930,000 Wikipedia articles.
The same application checks available memory on launch and downloads the model that fits; more memory can mean a smarter model. In the table below, 'Right' means the model got the main point of official guidance right on 22 blackout and first-aid questions. Results were graded blind in an emergency test on an earlier build of the library.
| Computer profile | Selected model | Emergency test | Model download |
|---|---|---|---|
| Mac with 128 GB or more memory | NVIDIA Nemotron 3.5 Super | 21 of 22 right | 80 GB |
| Mac with 32–96 GB | Qwen3.8 27B | 21 of 22 right | 18 GB |
| Windows PC with 32 GB or more | NVIDIA Nemotron 3.5 Lightning | 19 of 22 right | 25 GB |
| Mac or PC with 16–31 GB | IBM Granite 4.2 8B | 14 of 22 right | 6 GB |
Because smaller models make more mistakes, the interface always shows the source for manual checking.
The library download is 9 GB for all bundled sources, or 19 GB once installed. Without Wikipedia and iFixit, it is 1.3 GB downloaded and 3 GB installed. Free disk space should cover the model's size plus 5 GB, plus about 27 GB while the full library installs, or 10 GB for the smaller library. Compatibility is limited to Apple Silicon Macs (M1 or newer) running macOS 14 Sonoma or newer, and 64-bit Windows 10 or 11.
Setup depends on hardware and internet speed, from 20 minutes to about 2½ hours. Most clicks take about 5 minutes; the rest is download time.
On first open, the operating system warns that the app is not distributed through an app store. On Mac, click Done, open System Settings > Privacy & Security, scroll to the Survival LLM was blocked notice, click Open Anyway, and confirm with your password; this happens once. On Windows, click More info, then Run anyway.
The library holds plain-text copies of openly licensed references, including some that are for non-commercial use only, and public-domain materials. Search is local and fast.
Each article displays the license of its own source. That display is central to any open-source due-diligence workflow, because code, model, and reference text can carry different permissions.
AI models can make mistakes, and smaller models make them more often. The visible source list is the practical control: the user can check the original article before acting.
There is no account and no tracking. Questions remain on the local computer, and after setup the application can run with the internet off.
Users may add PDFs, Word documents, ebooks, and notes, such as a generator manual or a family emergency plan; the assistant searches those files too, and they stay on the computer.
At the bottom of the app, Check for updates installs a new version in one click while keeping models, the library, and user files.
On a MacBook Pro, measurement during answering was about 100 W, with each question taking 5 to 40 seconds; a full battery supports roughly 100 to 300 answers. It can run from a small generator or power station.