Oleksandr Kuzmenko, EditorPublished 6 min readAI-assisted · editor-reviewed
Why it matters
Engineers can deploy an immutable, offline reference architecture for critical operational manuals without cloud dependencies or data leaks.
TL;DR
- 01Survival LLM packages an offline RAG engine with curated medical, technical, and Wikipedia data.
- 02The installer dynamically provisions models from Granite 4.2 8B to Nemotron 3.5 Super based on RAM.
- 03MacBook Pro power consumption averages 100 W, supporting 100 to 300 answers on a single battery charge.
Key facts
- Full library download size
- 9 GB (19 GB installed)
- Minimal library download size
- 1.3 GB (3 GB installed)
- Power draw during inference
- ~100 W on MacBook Pro
- MacBook Pro battery yield
- 100–300 answers per charge
- Top accuracy benchmark
- 21/22 on blind blackout/first-aid tests
Hardware-Tiered Local Model Allocation
Survival LLM automates local inference configuration by matching installed memory with specialized weights. On systems with 16–31 GB RAM, the tool loads IBM Granite 4.2 8B (6 GB download, 14 of 22 correct in blind testing). Windows PCs with at least 32 GB utilize NVIDIA Nemotron 3.5 Lightning (25 GB download, 19 of 22 score). High-memory Apple Silicon systems receive Qwen3.8 27B (18 GB download, 21 of 22 score on 32–96 GB) or NVIDIA Nemotron 3.5 Super (80 GB download, 21 of 22 score on 128 GB+).
Offline Reference Index and Custom Ingestion
The full reference library requires a 9 GB download (expanding to 19 GB on disk) and incorporates MedlinePlus, CDC guidelines, military survival publications, Stack Exchange technical Q&A, and 930,000 top Wikipedia entries. A reduced 1.3 GB download (3 GB installed) strips Wikipedia and iFixit data. Users can append custom technical manuals, PDFs, and contingency documents directly to the local retrieval index without network interaction.
Power Consumption and Operational Realities
Benchmark measurements on a MacBook Pro indicate power draws of approximately 100 W during active generation runs lasting 5 to 40 seconds per question, yielding between 100 and 300 answers on a full battery charge. Because smaller models exhibit higher hallucination rates, the application pairs every generated completion with highlighted source passages from verified plain-text corpora.
✓ When to use
- Building offline field-ready workstations where internet access is unavailable or prohibited.
- Running private RAG over sensitive operational documentation without any telemetry.
- Prototyping local desktop model switching based on hardware RAM constraints.
✕ When NOT to use
- High-throughput multi-user API serving across production web applications.
- Low-spec machines with under 16 GB RAM or insufficient disk space (requires model size plus 27 GB free).
- Tasks demanding real-time updated web data or dynamic cloud API calls.
What to do today
- Verify system resources: ensure at least 16 GB RAM and 30+ GB free storage before installation.
- Download the standalone archive from brangus.io to prepare an offline air-gapped recovery node.
- Index custom disaster recovery and hardware runbooks using the integrated local document loader.
What the community says
“This could be made a lot more practical (and safer) if made searchable via a simple vector search / embedding that can run on any ordinary computer.”
“Granite is not great at any size. Wonder why they picked that.”
“How on earth are you going to operate this if you have no electricity?”