Survival LLM answers emergency survival questions without internet

Summary

Survival LLM is a free, fully offline AI assistant (v0.3.4, October 2026). It answers first aid and survival questions by querying a locally stored library of approximately 930,000 Wikipedia entries, FEMA/CDC guidelines, and military field manuals, citing sources for every response. It adjusts the model size based on the host machine's RAM, ranging from a 6 GB IBM Granite model to an 80 GB NVIDIA Nemotron model.

Why it matters

As cloud-based AI becomes the default, a fully offline emergency tool addresses a critical resilience gap. When power grids and networks fail, hosted large language models (LLMs) become unavailable. Thanks to its source-citation design and the ability for users to add documents (generator manuals, family plans), this tool serves as a practical complement to traditional paper-based emergency kits rather than a replacement.

Highlights

• Runs 100% locally with zero external requests; every response includes a link to a highlighted source passage for verification
• Automatically selects the model based on available RAM: from 6 GB (for 16-31 GB machines) to 80 GB (for 128 GB+ Macs); the library can shrink from 9 GB down to 1.3 GB
• In a 22-question test, the modest-hardware model answered only 14/22 correctly; on a MacBook Pro, each query consumes approximately 100 W for 5-40 seconds

Critical take

A 36% error rate on the most accessible hardware tier poses a serious risk, especially given the tool's positioning for life-or-death first aid decisions. Additionally, the approximately 100 W energy consumption per query undermines the 'power outage' scenario the tool targets. While the project is a thoughtful proof of concept, claiming it is 'designed for emergencies' is an overstatement beyond what a 14/22 accuracy score can responsibly support.