Survival LLM Scores 14 of 22 on Laptops With 16 GB
Software / news
Survival LLM Scores 14 of 22 on Laptops With 16 GB
The free offline emergency assistant at version 0.3.4 grades its own tests, names no developer, and links to a placeholder for its source code.
Survival LLM, a free MIT-licensed app that bundles a language model with an offline reference library, reached version 0.3.4 on Oct. 27, and its own test of 22 emergency questions shows the small-laptop model getting 14 of them right. The project page names no developer, and the link it gives for the source code is a placeholder.
The app runs on Apple Silicon Macs with macOS 14 or newer and on 64-bit Windows 10 or 11. On first launch it checks installed memory and downloads whichever model fits, then pulls the library. Nothing a user asks leaves the machine once setup finishes, the page says, and there is no account.
Which model a laptop gets, and how it scored
The page lists four tiers. The scores are the developer's own: 22 blackout and first-aid questions, graded blind, run on an earlier build of the library. No outside party is named as having repeated them.
| Computer | Model | Score | Download |
|---|---|---|---|
| Mac, 128 GB or more | NVIDIA Nemotron 3.5 Super | 21 of 22 | 80 GB |
| Mac, 32 to 96 GB | Qwen3.8 27B | 21 of 22 | 18 GB |
| Windows, 32 GB or more | NVIDIA Nemotron 3.5 Lightning | 19 of 22 | 25 GB |
| Mac or PC, 16 to 31 GB | IBM Granite 4.2 8B | 14 of 22 | 6 GB |
- Nemotron 3.5 Super21 of 22
- Qwen3.8 27B21 of 22
- Nemotron 3.5 Lightning19 of 22
- Granite 4.2 8B14 of 22
Source: Survival LLM project page (developer's own test, earlier library build), accessed 2026-10-29
The sticking point is the bottom row. A machine with 16 GB of memory gets the model that missed eight of 22. The page says plainly that "AI models make mistakes, the small ones more often," and tells users to open the cited article behind each answer. That design choice matters more than the score: each reply lists the library articles it drew from, with the relevant passage marked.
What the library holds
The library is plain-text copies of openly licensed or public-domain references. The page lists about 930,000 of Wikipedia's most-read articles, WikiMed, MedlinePlus, WikEM, Ready.gov guides from FEMA, CDC travelers' health pages, U.S. military medicine and survival manuals, iFixit repair guides and Stack Exchange boards on home repair, gardening and ham radio.
The full library is a 9 GB download and takes 19 GB installed. A trimmed version without Wikipedia and iFixit is 1.3 GB and takes 3 GB. Free disk space needed is the model size plus 5 GB plus about 27 GB while the full library installs. Setup takes 20 minutes to about two and a half hours, mostly downloading.
Each source keeps its own licence, and the page warns that some allow non-commercial use only.
Power draw is the number to check
The page puts a MacBook Pro at about 100 watts while answering, for 5 to 40 seconds a question. A full battery handles roughly 100 to 300 answers. That is the figure a person in a blackout needs, and it is the developer's estimate, not a measurement anyone else has published.
The models themselves are third-party. Qwen3.8 27B is Apache 2.0 with a 262,144-token native context, according to its Hugging Face card, which gives no hardware requirements. That means the 18 GB download figure and the 32 GB memory tier come from Survival LLM, not from the Qwen team.
What it does not tell you
The page names no maintainer, the code link is a placeholder, and the first-run screen triggers an operating system warning because the app is not sold through an app store. The page tells Mac users to open System Settings, then Privacy & Security, then Open Anyway.
Compare that with JetBrains' ThinkRail, whose telemetry default is covered here, and with Recursive's Flint coding model, which carries an Apache licence. Survival LLM promises no tracking and names no one to ask.
The next check is the repository link. If the source code appears under the MIT licence the page claims, the 22-question test set can be rerun by someone else. Until then the scores are the developer's alone.
Sources
More in Software
- 01Python 3.15 Ships a Built-In sentinel After Five Years of PEP 661The new type replaces the object() idiom with a one-line constructor that survives copying, pickling and type checking, but a Python 3.15 floor is the price of using it.
- 02Lightpanda 1.0 Passes 80% of Chrome's Web Tests, by Its Own CountThe Zig-built headless browser left beta on Oct. 2, and its own benchmarks put it ahead of two Rust rivals on a 940-page crawl, with no hardware listed.
- 03Claude-Mem Hits 107,900 Stars, and Its Own Pages Disagree on LicenceThe Claude Code memory plugin has gained about 35,500 stars since May, but a May write-up says AGPL-3.0, the repository now says Apache-2.0, and the README describes an optional cloud sync that earlier coverage did not.
- 04Jepsen Found 96 Issues in ParticleDB 1.1.4 and Tells Users to Halt DeploymentKyle Kingsbury’s 31 October report, funded by vendor Orkes, finds lost writes and ignored constraints in healthy single-node runs.