Migrating off a local AI tool is usually an archaeology problem. You go looking for your models and find a content-addressed store, a manifest layer, and a community script from 2024 that used to work.

LM Studio is not that. It writes plain GGUF files with the names they were published under, in a directory it lets you change. Whatever else is true about the app, its storage is honest, and the consequence is that this guide is short.

Where are the models?

By default, ~/.lmstudio/models, organized into publisher and repository subfolders that mirror how the models are laid out on Hugging Face. The folder is user-changeable from within the app, and LM Studio records the choice in its settings, so if you ever pointed it at an external drive, check the setting before assuming the default.

The fastest way to confirm what you actually have is the CLI. LM Studio ships lms with the app, and lms ls lists the models available on disk. Run it before you move anything and you have your inventory. If the command is not found, the app has not been run yet on this machine, which the docs list as a prerequisite.

Moving to another llama.cpp-based tool

Copy or move the .gguf files into the new tool's models directory. There is no conversion, no import step, and no metadata to rebuild. GGUF carries its own architecture, quantization and chat template in the file header, so the receiving tool reads everything it needs from the bytes you moved.

Two practical notes. Copy rather than move while you are testing, so a bad result costs you disk rather than a redownload; delete the originals once the new tool has actually generated something. And keep the filenames, because they encode the model name and quantization, and you will want to know which is which in six months.

Moving to Ollama is the one direction that is harder

This is the query people search most in this cluster, and it deserves the honest answer: it is the awkward direction, and not because of anything LM Studio did.

Ollama does not read loose GGUF files from a folder. It keeps weights in a content-addressed store as sha256-named blobs behind its own manifest format, so importing means writing a one-line Modelfile and running ollama create, which copies the file into that store. Your disk usage roughly doubles until you delete the original, and the model is then inside a layout that takes manual digest mapping to get back out of. Importing into Ollama is a one-line Modelfile; exporting from it has no command at all.

If Ollama is where you are going, do it deliberately and one model at a time, watching the disk. If you are going there because you want a headless server, that is a good reason. If you are going there because you assumed it was the standard destination, know that any llama.cpp-based tool takes your files as they are.

What does not transfer

Honesty section. Three things stay behind, and none of them are recoverable from the model files.

Chats. Conversation history is app state, stored in LM Studio's own folders, in a layout that is LM Studio's and version-volatile. There is no portable format for this across local AI apps, so treat the history as something you leave rather than something you move. If a specific conversation matters, copy the text out before you uninstall.

Per-model presets. LM Studio lets you save defaults per model, including GPU offload, context size and whether to use Flash Attention. Those live in app state, not in the GGUF, so you will set them again wherever you land. Most of them you will not want anyway: GPU offload on Apple Silicon should be at maximum in any tool, and context size is a decision you should be remaking against the new tool's memory behaviour.

Prompt templates you overrode. If you edited a model's template in the app, the edit is app state. The template that travels is the one embedded in the GGUF metadata by whoever published it, which is usually the correct one and is what the receiving tool will use.

Doing it, step by step

The steps below move one model so you can verify before committing. Repeat for the rest, then reclaim the space.

Landing in ToolPiper

You do not move anything into ToolPiper. It finds LM Studio's library on its own, reading the same home pointer and models-folder setting LM Studio uses, so a relocated folder is found too, and asks once to adopt it, with the model count and total size on the button. Adoption hardlinks each GGUF into ToolPiper's own store: no bytes are copied, nothing is written into LM Studio's folder, and the adopted models keep working if you later uninstall LM Studio. When a hardlink is not possible, ToolPiper offers a real copy instead. One side effect to know: while both apps hold a link to the same file, deleting the model in LM Studio does not free the disk space. LM Studio's per-model presets are not read.

The engine is upstream llama-server embedded directly, build b12359, unmodified, so the same file behaves the same way. Expect parity on speed rather than a gain: our same-bytes benchmark on an M2 Max landed within single digits in both directions with the winner flipping by model. Pick on storage, interface and direction.

What the free tier adds around the runner: unlimited model downloads, multi-model switching, a local OpenAI-compatible API, embeddings, developer tokens, all speech including transcription, text-to-speech, voice cloning and push-to-talk dictation, Apple Intelligence, and an MCP host that Claude Code, Cursor and any other MCP client can call, with 207 of ToolPiper's 470 native tools free. No account, no caps, no terminal. Pro ($10/month) is our tools: browser automation, files and Git, the clipboard cache, local RAG over your files, web scraping with YouTube transcripts, and the cloud API proxy for your own keys.

ToolPiper runs on macOS 26 or newer on Apple Silicon. If in-app model discovery is what you use LM Studio for, keeping it installed alongside works well, because the files are shared either way.

Cost, since a migration is a good moment to check it: LM Studio's local app is free, and its cloud models are sold through the Bionic+ ($20 a month) and Pro ($100 a month) plans. Nothing in the move above costs anything on our side either, and the free tier covers the runner outright. The full comparison of what each charges for.