What runs

Software

Failsafe runs an open model on your device, with a library you can read. The file it uses depends on the memory on that device.

This website does not run the AI. The details below come from the current prototype: what is installed for testing, and what is only listed in the catalog. Device testing is still pending. This page does not publish benchmark scores.

How a model is chosen

A picker selects a catalogued model that fits the device. A smaller device uses a smaller helper. A device with more memory can use a larger catalogued file. Those pairings are planning guides from the prototype catalog, not speed scores from Failsafe device testing.

Answers can use the model’s trained knowledge and a downloaded reference library. That is not live news. Official model cards for the models named here do not publish a knowledge-cutoff year, so this page does not invent one.

Engine

Used by the prototype

llama.cpp llama-server, MIT license, pin b12167. The prototype launcher starts it on this computer and serves a local interface. It does not send your questions to a cloud API.

Ask models

Official Apache-2.0 Qwen3 GGUFs from Qwen (Alibaba Cloud). Ask mode can search a local library and cite the pages it used.

  • Installed in the prototype

    Qwen3 0.6B Q8_0

    File Qwen3-0.6B-Q8_0.gguf. Present in the research prototype used to test Failsafe. Official card: Qwen/Qwen3-0.6B.

  • Catalogued, not installed here

    Qwen3 4B, 8B, 14B, and 32B Q4_K_M

    Listed for computers with more memory. They are not present in this research folder.

Coding models

Official Qwen2.5-Coder Instruct GGUFs. Coding mode formats answers as code and skips the household library.

  • Installed in the prototype

    Qwen2.5-Coder 1.5B Instruct Q4_K_M

    File qwen2.5-coder-1.5b-instruct-q4_k_m.gguf. Present in the research prototype. Apache-2.0. Official card: Qwen/Qwen2.5-Coder-1.5B-Instruct.

  • Catalogued, not installed here

    Qwen2.5-Coder 7B, 14B, and 32B Instruct

    Apache-2.0 files listed for higher-memory machines. They are not present in this research folder.

  • Listed, never auto-selected

    Qwen2.5-Coder 3B Instruct

    Noted in the catalog for completeness. It uses the Qwen Research License, not Apache-2.0, and is never chosen automatically.

What this prototype does

It answers questions from the model and a saved library, and it can help with writing and coding. It is local Q&A on your device, not a system that acts on the device for you.