What actually ships§
Survival LLM, documented at brangus.io/survival-llm, is a desktop application that installs a language model and a document corpus onto the user's own machine. Setup requires an internet connection; after that, the project states it needs none. Nothing the user asks leaves the computer.
The corpus is the interesting part. It is not a generic web crawl but a curated set aimed at emergencies: medical references, FEMA and CDC guidance, survival manuals, how-to and repair guides (iFixit is bundled), and approximately 930,000 Wikipedia articles. The full library is a 9 GB download that expands to 19 GB installed; a reduced version without Wikipedia and iFixit is 1.3 GB down and 3 GB installed.
The interface is deliberately auditable. Every answer lists the articles it consulted, and clicking one opens the source with the exact passage highlighted. That design choice is the product's core argument: in a blackout or a first-aid scenario, an unverifiable answer is worse than no answer, so the citation is not a nicety but the safety mechanism.
The architecture: retrieval over a memory-scaled model§
The described behavior maps closely onto retrieval-augmented generation. Rather than expecting a model to memorize water-purification ratios or burn-treatment steps — which small models reliably garble — the system retrieves relevant passages from the installed corpus and conditions the answer on them, then exposes those passages to the user. The retrieval layer also explains why the library can be split: dropping Wikipedia and iFixit from 19 GB to 3 GB removes breadth, not the official guidance that most emergency answers depend on.
The model tier is chosen at first launch. The app inspects the host machine's memory and downloads the model that fits, with more memory yielding a smarter model. This is a pragmatic concession to the reality of consumer hardware: a single shipped model would either exclude low-RAM machines or bore high-RAM ones. But it also means answer quality is a function of the user's hardware, not a fixed property of the product. Two people running Survival LLM may be running materially different systems.
Platform support is narrow and modern: Apple Silicon Macs (M1 or newer) on macOS 14 Sonoma or later, and 64-bit Windows 10 or 11. The absence of Intel Macs and older Windows is consistent with a stack that assumes accelerated local inference. Disk budgeting is explicit — model size plus 5 GB free, and roughly 27 GB of transient space during a full library install (10 GB for the reduced one). Setup time ranges from about 20 minutes to 2.5 hours depending on the machine, a range wide enough to signal that install is doing real work: downloading tens of gigabytes and likely compiling or warming model artifacts.
The 22-question claim, read carefully§
The project's headline correctness statement deserves parsing. "Right" is defined as getting the main point of official guidance correct on 22 blackout and first-aid questions, graded blind, in the project's own tests, and — critically — on an earlier build of the library. The project itself notes that smaller models make more mistakes, which is why the app always shows the source.
That is an honest framing, and it is also a thin one. Twenty-two questions is a smoke test, not an evaluation suite; it cannot separate a 7B-class model from a 30B-class one with any statistical confidence, and the per-model breakdown is not published in the excerpt. Grading was blind, but the question set was presumably authored by the same team that built the retrieval corpus, which raises the familiar risk of a test aligned to the library's strengths. The caveat that results came from an earlier library build means the current release is, strictly speaking, unmeasured. None of this invalidates the approach — the always-visible citation is a better safety property than a benchmark score — but practitioners should treat "22/22" as a claim about methodology rather than a capability ceiling.
Practical implications§
The project is best understood as a reference implementation of a pattern rather than a novelty. It demonstrates that a credible offline assistant is an integration problem — corpus curation, chunking and retrieval quality, citation rendering, model-tier negotiation, and install-time disk management — as much as a modeling one. Teams building field tools, clinical decision support in low-connectivity settings, or air-gapped enterprise assistants face the same trade-offs at larger scale.
The constraints are real. A 19 GB install is not trivial on a machine that may also need to hold your documents, and the 27 GB transient requirement during install will fail on drives with less headroom. Hardware gating excludes a large installed base. And the corpus, while curated, is frozen at install time — medical guidance and FEMA documentation get revised, and an offline library has no mechanism described here for surfacing that its sources have aged.
What remains uncertain: which model families and parameter counts the app selects per memory tier, the retrieval method and chunking strategy, the license terms of the bundled corpus components, and whether the project intends to publish a larger, versioned evaluation. Until then, Survival LLM is a well-designed argument for local, cited AI — with a test suite too small to settle the question it raises.