For three years, "AI" meant somebody else's data center. October 2026 is when that stopped being true. VoxNexus showed its VoxEdge stack running a complete voice pipeline — wake word, speech recognition, understanding, spoken reply — entirely offline on inexpensive Android phones. OpenAI added PIN locks to ChatGPT, a small feature with a big admission: people want AI that answers to them, not to a server. The reasons are stacking up. Privacy: your prompts, documents and recordings never leave the device. Cost: once the model is downloaded, the per-token bill is zero, forever. And reach: offline AI works on a plane, in a rural clinic, in a classroom with no signal — anywhere the cloud simply isn't. This hub collects the tools that make it real.
Every serious way to run AI without the internet, in one place. All of these are free to start with; the models they run are open weights you download once and keep.
Run open-weight models locally with one command; backed by a huge community model library with hundreds of ready-to-pull models and quantizations.
A polished desktop app for downloading, managing and chatting with local models — no command line required, with a built-in model browser and local API server.
The C/C++ inference engine most local tools are built on. The GGUF quantization formats that make big models fit small machines come from this project.
An open-source, privacy-first desktop AI assistant designed from the start to run fully offline, with your conversations stored only on your machine.
Nomic's local LLM app, built to run on ordinary CPUs — no discrete GPU required. Includes a local document-chat feature that keeps files on-device.
An on-device voice stack unveiled in October 2026. The vendor claims fully offline speech recognition in 36 languages running on cheap Android phones — vendor claims, not independently verified by us yet.
High-throughput serving for self-hosted models, using PagedAttention to squeeze far more concurrency out of the same GPU. This is the step up when "local" grows into a server for a team.
Three steps between you and a model that answers with the Wi-Fi off:
Yes. Open-weight models run entirely on your own hardware — laptop, desktop, server, or even a phone — using runtimes like Ollama, LM Studio or llama.cpp. Once the model file is downloaded, no prompt, document or recording ever has to leave the device. Cloud AI is a hosting choice, not a technical requirement.
Less than most people expect. Small quantized models (1B–8B parameters) run on ordinary laptops and modern phones; a 7B model at 4-bit quantization needs roughly 4–6GB of free RAM. Bigger models want a GPU with more VRAM. Our Can I Run It? calculator matches a model to your exact hardware before you download anything.
The frontier cloud models are still ahead on the hardest reasoning and coding tasks. But for summarization, drafting, transcription, translation and private document work, current open-weight models running locally are genuinely good — and they are free per token, private by design, and available on a plane, in a clinic, or anywhere with no signal.
The software in this directory is free and mostly open source, and the models are free to download. Your only costs are the hardware you already own and the electricity it uses. There is no per-token bill, no subscription, and no usage cap.