Brangus Labs just released Survival LLM, a downloadable package that combines a local language model with offline reference libraries aimed at disaster preparedness. It’s exactly what it sounds like: AI assistance that works when your internet doesn’t.

The project packages several open-weight models (Llama 3.2, Phi-3.5, Qwen2.5) alongside compressed Wikipedia archives, survival guides, medical references, and other emergency documentation. Everything runs locally on your computer. No API calls, no cloud dependency, no data leaving your machine.

You can query the embedded knowledge base through a chat interface, asking questions about first aid, water purification, radio operation, whatever’s in the bundled docs. The models aren’t fine-tuned for survival scenarios, they’re just standard instruction-following LLMs with RAG over the reference material.

What’s actually in it

The knowledge base includes Kiwix archives (compressed Wikipedia and other wikis), Project Gutenberg books, US military field manuals, ham radio licensing guides, and medical references. Total download is several gigabytes depending on which package you grab.

The interface is straightforward: local web UI, chat with the model, search the docs directly if you want. It uses Ollama for model inference, so if you’re already running local models, setup is trivial.

Hardware requirements are reasonable. The lighter models (Phi-3.5-mini) run fine on modest hardware. The larger ones need more RAM and benefit from a GPU, but nothing exotic.

Who this is for

This makes sense if you’re into disaster prep and already comfortable running software from GitHub. It’s not a consumer product. You need to know how to install dependencies, run terminal commands, and troubleshoot when things break.

The use case is genuinely niche: extended internet outages, remote locations, situations where you can’t or won’t connect to external services. If you’re building an off-grid setup or just want your AI tools to work during network failures, this solves a real problem.

Privacy-focused users might also care. Everything stays on your machine. If you’re working with sensitive information and don’t want it hitting external APIs, local inference is the only option.

The limitations

Don’t expect GPT-6 performance. These are smaller open-weight models running on consumer hardware. They’re competent for basic question-answering over the knowledge base, but they’re not doing complex reasoning or creative tasks well.

The knowledge base is static. You download it once, it doesn’t update unless you manually pull new archives. Wikipedia content gets stale, manuals get revised, medical information changes. This isn’t a living reference.

And obviously, if the scenario you’re prepping for involves losing power, you need to solve that separately. This requires a working computer.

Why it exists

The creator frames this as a response to increasing reliance on cloud services that might not be available in emergencies. It’s a valid concern. When Hurricane Helene knocked out cell service and internet across parts of North Carolina last month, people couldn’t access basic information they’d normally Google.

The project is open source (MIT license) and actively maintained. The repo includes documentation for adding your own reference material and swapping models.

This won’t replace your emergency kit, but if you’re the kind of person who keeps offline map tiles and downloaded documentation, Survival LLM fits the same mindset. It’s a specific tool for a specific problem.

The code is on GitHub. If you’re curious, clone it and try the lighter models first. If you’re not already thinking about offline computing, you probably don’t need this.