LM Studio app: the cross-platform desktop application

UI tour: the four main tabs

Four tabs in the left navigation rail cover every core workflow: Chat for inference, Discover for model management, Server for API access, and Settings for hardware tuning. Each opens a dedicated pane; switching tabs does not unload the active model.

The Chat tab in depth

The Chat tab is where most users spend the majority of their time in the LM Studio app. It combines a system prompt editor, a sampling preset selector, and a conversation thread in a single scrollable view.

At the top of the Chat pane sits the model selector — a dropdown that lists all downloaded models. Switching models here unloads the current one and loads the new selection; a status bar at the bottom tracks progress and shows the resulting token rate and memory footprint once loading completes. Below the selector is the system prompt field, which accepts arbitrary text and is included in every request as the first message in the messages array. Changes to the system prompt take effect from the next user turn, not retroactively.

The right-hand panel contains the sampling preset. Temperature, top-p, top-k, repeat penalty, and max-tokens are the controls most users adjust. LM Studio ships with several named presets — Precise, Balanced, Creative — that set sensible values for different output styles. Custom presets can be saved and recalled, which is useful when you need to switch between a deterministic code-generation setup and a more exploratory creative writing configuration without manually re-entering values each time.

Conversations are automatically saved to disk as they accumulate. The history sidebar on the left of the Chat pane shows previous sessions, each labelled with the first user message as a title. Clicking a past session reloads its full transcript. The export button at the top of any session writes the conversation to a Markdown or JSON file in a location you choose.

The Discover tab in depth

The Discover tab is the LM Studio app's model marketplace. It connects to a curated index of Hugging Face repositories and presents quantized GGUF variants alongside hardware-fit ratings and file sizes.

Searching the Discover tab returns results from a manifest that the LM Studio app syncs from the network at startup. If you are offline, the search field still works against the local cache of the manifest from the last sync. The manifest includes model metadata: parameter count, context length ceiling, architecture family, and the list of quantization variants with their corresponding download URLs and file sizes.

When you download a model, the file lands in your configured models directory. A progress bar in the Discover tab tracks the transfer. Downloads are resumable: if the application is closed during a download, LM Studio picks up where it left off the next time you open the Discover tab with a network connection. Completed models move from the search results list to a separate Downloaded section visible at the top of the Discover pane.

Supported file formats

The LM Studio app loads GGUF files exclusively. This single-file format bundles weights, tokenizer configuration, and model metadata, removing the need for separate config downloads.

GGUF is the current standard for llama.cpp-compatible quantized models. A file ending in .gguf with a valid GGUF header loads in LM Studio without additional steps. Split GGUF shards — named like model-00001-of-00003.gguf — are also supported: place all parts in the same folder and open the first shard. LM Studio stitches the pieces together at load time.

The older GGML format is not supported in recent LM Studio versions. If you have GGML files from an earlier workflow, convert them to GGUF using the llama.cpp conversion utilities before attempting to load them. SafeTensors and HuggingFace bin files (the native HuggingFace format) are also not directly loadable — they must be converted to GGUF first, a process that typically involves quantization and takes a few minutes on a reasonably fast machine.

Configuration directory and storage locations

The LM Studio app follows platform conventions for storing data. Knowing the config directory lets you back up models, migrate to a new machine, or script around the application’s file structure.

On macOS, the LM Studio app stores its configuration and data at ~/Library/Application Support/LM Studio/. Model files live in ~/Library/Application Support/LM Studio/models/ by default, though this path is settable in the Settings tab. Conversations are stored as JSON in ~/Library/Application Support/LM Studio/chats/. Application logs for troubleshooting are in ~/Library/Logs/LM Studio/.

On Windows, the base directory is %APPDATA%\LM Studio\ (typically C:\Users\YourName\AppData\Roaming\LM Studio\). Models go in %APPDATA%\LM Studio\models\ by default. Chats are at %APPDATA%\LM Studio\chats\.

On Linux, the LM Studio app uses ~/.cache/lm-studio/ as the base. Model files land in ~/.cache/lm-studio/models/ and conversations in ~/.cache/lm-studio/chats/. The application itself is distributed as an AppImage and does not install to system directories, making it straightforward to run multiple versions side-by-side by keeping them in separate folders.

To migrate LM Studio to a new machine, copy the models directory to the same path on the new machine (or point Settings at its new location) and the app recognises all previously downloaded models immediately. Conversation history can be migrated the same way by copying the chats directory. See NIST’s AI Risk Management Framework and FTC guidance on AI products for context on data handling when deploying AI applications.

How conversations are stored

Each conversation in the LM Studio app is saved as an individual JSON file on disk, labelled with a timestamp. The file structure is human-readable and straightforward to parse for custom tooling or archival workflows.

Every time you start a new chat session, LM Studio creates a new JSON file in the chats directory. The file accumulates messages as the session progresses — each user and assistant turn is appended as an object with role, content, and a Unix timestamp. The session file is written to disk after each turn, so a crash does not lose the current conversation beyond the last completed exchange.

The JSON schema mirrors the OpenAI messages format, which means conversation files can be loaded directly into scripts that use the OpenAI SDK — parse the JSON, extract the messages array, and pass it as the messages parameter in a chat completion call. This makes it trivial to resume a conversation programmatically, replay it against a different model, or build tooling that analyses session history. The LM Studio app also provides a one-click export from the Chat tab UI that writes the session to a clean Markdown file, which is more readable for archival and handoff than raw JSON.