The dial draws the Mac's unified memory as a 270° ring. Each loaded model is an arc in its own colour. macOS and other apps are a dim arc, and the 15% safety reserve is hatched. The empty track that remains is the memory a new model can take, and the figure in the centre is its size. Point at an arc or at its legend row to read that segment's name and size.

Three tiles under the legend show CPU load over the last minute, free disk space in the models folder, and swap. When swap passes 1 GB, the tile turns amber and a warning suggests stopping a model or restarting it with a shorter context, because generation slows down once macOS pages memory to disk. Each running model has Chat, Use from code and Stop, and a bar with its share of memory in the colour of its arc. The activity column lists downloads, starts and stops.

The core is a ring around a disc. It appears in the rail, in the model lists, in the command palette, in the chat model picker and next to every reply. The six states change how much of the ring is lit, how full the disc is, and whether anything moves. Only loading and generating animate continuously. A running model is drawn still, because a loaded model that is not generating does no work.

A transition to running plays one arrival, in which the disc overshoots and a ripple leaves the ring. A transition to failed shakes the core once. Neither plays when a view first opens, only on a change of state. The buttons below drive one core through a load.

The search runs against the Hugging Face Hub 300 ms after you stop typing. Filters limit the results to 4-bit or 8-bit weights, vision models, instruct models, or models that fit in free memory now. Results sort by downloads in the last 30 days, by likes or by last update, 30 at a time. ⌘F focuses the search, and the command palette can send what you typed there.

Each row has a fit gauge. The track is what the Mac can give a model at all, which is total memory minus the reserve. The bar is what the model needs, and the white notch marks what is free right now. A model that is not downloaded has no config.json to read yet, so its estimate is the weight size from the parameter count and the quantization, plus 10% for the KV cache and 15% for runtime overhead.

Start opens a dialog with a context-length slider in steps of 1,024 tokens, up to the model's max_position_embeddings. The estimate is recomputed 180 ms after the slider stops. For an installed model, the weights are the size of the safetensors files on disk, and the KV cache comes from the model's config.json.

The strip under the figure is the dial's ledger, one cell per GB. The model you are about to start is drawn as outlined cells in the colour it will get. Memory it needs beyond what the Mac has is drawn as red cells past the end of the strip. Start is disabled only when the model is too big at the chosen context. The Reasoning switch sets enable_thinking, for models such as Qwen3 and DeepSeek R1 that reason before they answer.

A download lists the repo's files on the Hub, keeps the ones needed to run the model, and fetches them one at a time. Progress and speed reach the app over SSE, and the speed is smoothed over samples of the size on disk. Pause and Cancel are checked between files, so a pause takes effect when the current file finishes. On resume or retry, files that are already complete on disk are skipped.

A second Download of the same repo returns the job that is already running, because two workers writing one folder would corrupt it. A failed job keeps its files and retries from its row. Cancel asks first and says how many bytes it deletes. Clear finished removes the rows of completed jobs and leaves the installed models alone. The badge on Downloads in the rail counts the jobs that are downloading or queued.

Running models sit in a lit group above the installed ones. Start and Stop are always visible. Use from code, the model log, Check for updates and Delete appear on the row under the pointer, so a list of ten models stays a list of ten names. Delete asks first, names the repo and the bytes it frees, and unloads a running model before it removes the weights.

A repo with no chat template, such as Whisper or an embedding model, is marked Not a chat model. It has no Start button, and neither Chat nor the API can use it. The log dialog follows the model's log live: context length and estimated memory at load, load time, prompt speed and peak memory per generation.

A result that lands outside the view you are on, such as a model finishing its load, a delete or a failure, appears as a toast in the bottom-right corner. Errors stay on screen longer than successes.

The picker in the header lists the running models. A message goes to the same /v1/chat/completions endpoint external clients use, and the reply is rendered once per animation frame while it streams. A model that reasons first, such as Qwen3 or DeepSeek R1, shows a Thinking label while it reasons. The reasoning then folds into a Thought for N s control above the answer.

Code blocks are highlighted and each has a copy button. A finished reply shows copy, regenerate, tokens per second and time to first token. When a reply stops at the token limit, a notice under it says so and names the Longest reply setting. With a vision model selected, the composer accepts up to four images from the picker, a paste or a drop.

Conversations are grouped by date, filtered by title and renamed from the header. The system prompt, the temperature and the longest reply apply to every chat and are kept across launches.

The sidecar serves /v1/models and /v1/chat/completions on 127.0.0.1:11535. /v1/models lists every installed chat model. A request for an installed model that is not running loads it first and then answers, so a client never has to start a model through the app. Requests are checked against the API key shown in Settings. The app's own calls to the rest of the sidecar use a separate token that the Rust core generates at each launch.

Tool calling follows the OpenAI format: tools, tool_calls, messages with the tool role, streamed deltas and finish_reason: "tool_calls". It works with models whose chat template declares a tool format that mlx-lm can parse, such as Qwen, Mistral, GLM, Gemma and Kimi. A request with tools to any other model returns 400. Vision models accept image_url content parts, as base64 data URLs or http URLs.

Generation runs on a single MLX thread, so two requests that arrive together are answered one after the other. Use from code opens the dialog below with the base URL, the key, the model id and a working client in Python, LangChain, Java with LangChain4j, Rust with Rig, or curl. The dialog remembers the last language you chose.

The palette lists the six pages, the commands for each model and the app actions. A running model offers Chat and Stop. An idle model offers Start, which opens the Start dialog. Matching is fuzzy and ranks matches at the start of a word first. Whatever you type can also be sent to the catalog search. The arrow keys move, Enter runs the selected command, and Escape closes the palette.

The menu-bar item shows the memory free for models and a submenu for each running model with Open chat and Stop. It reads /system/stats and /models from the sidecar every 3 seconds. Closing the window hides it and leaves the engine running. The Dock icon or Open MLX Studio brings the window back. Quit MLX Studio or ⌘Q stops the engine and exits.

Appearance follows the system or forces light or dark. The light theme keeps the layout on a paper-coloured canvas and uses the lime only as a fill, never as text. Chat holds the defaults described above. Local API shows the base URL and the key to copy. A Hugging Face token is optional. It raises the Hub's rate limits and gives access to gated repos. Storage shows the models folder, and Memory draws the fit rule that the catalog and the Start dialog apply.

MLX, mlx-lm and huggingface_hub are Python libraries, so the engine runs in a Python sidecar. The shell is Tauri, which uses the system WebView instead of bundling Chromium. In a packaged build the sidecar is a PyInstaller binary inside the bundle, so nothing has to be installed first: no Python and no Homebrew. In development the three run as separate processes, and the frontend falls back to localhost defaults when it runs in a plain browser.