LM Studio

LM Studio is a free desktop application for discovering, downloading, and running open-weight large language models entirely on your own computer. It targets users who want local AI without living in a terminal, pairing a model browser, a chat interface, and a local API server in one application. For many sovereign Bitcoiners it is the most approachable on-ramp to running a model offline: install it, download a model, and you are having a private conversation with hardware you own within minutes.

What it offers

LM Studio is built on the llama.cpp engine and runs on macOS (Apple Silicon), Windows, and Linux. It loads models in GGUF format and, on Apple Silicon, also supports Apple's MLX backend for faster inference on that hardware. The built-in browser searches the Hugging Face Hub and flags which quantization of a model will fit in your machine's memory — genuinely useful guidance, since choosing among a dozen GGUF variants is the first wall a newcomer hits. Beyond the chat UI it includes document chat for retrieval over your own files, an lms command-line tool for scripted workflows, and a headless service mode for running on a server without the GUI. GPU offload is adjustable per model, so a machine with modest VRAM can split layers between GPU and CPU rather than giving up.

Local API compatibility

Like other tools in this space, LM Studio exposes an OpenAI-compatible API server on localhost (port 1234 by default), so applications written for hosted AI services can be pointed at your private model by changing one base URL. That makes it a practical drop-in backend for editors, agents, and scripts — the pattern that matters for sovereignty, because it means the ecosystem of AI-powered tooling can run against a model that never sends your data anywhere. A machine that already lives in your workshop — including one whose waste heat is doing useful work — can serve the whole household's AI over the LAN.

The licensing caveat

Note that LM Studio itself is a proprietary application rather than open source, though the engine (llama.cpp) and the model formats it relies on are open. Nothing about your prompts or documents leaves the machine during inference, but a closed binary is still a trust decision, and the verify-don't-trust instinct applies to AI tooling as much as to wallets. For a fully open-source workflow, the underlying components remain available directly — llama.cpp from a terminal, or open server frontends — at the cost of more setup. A reasonable path: start with LM Studio to learn what local models can do, then migrate the pieces you depend on to fully auditable tooling as your stack matures.

Where it fits

LM Studio occupies the "approachable desktop" tier of the local-AI stack: heavier than a bare llama.cpp binary, friendlier than a self-hosted server rig, and a solid daily driver for private chat, document Q&A, and serving one machine's models to your own applications. It runs the same GGUF models as the rest of the ecosystem, so nothing you download is locked in — the models remain yours, on disk, whatever frontend you choose tomorrow.

Sizing hardware is mostly a memory question: the model file must fit in RAM (or VRAM for full GPU offload) with headroom for context, so the practical ceiling on a given machine is set by how much memory it has and which quantization you choose — smaller quantizations trade a little quality for a lot of headroom. Start with a mid-size model that fits comfortably rather than the largest that technically loads; a responsive assistant you actually use beats an impressive one that swaps.

Compare local-AI apps in the sovereign self-hosting catalog.