New AI Workflows • Cut Out the Subject • Upscale a Small Photo • A Voiceover Bed Local Server, now with its own screen • Turn it on in Settings → Local Server → Enable Local Server. A new Local Server screen in the sidebar shows what you’re serving, live status, memory use, request stats and a recent-requests log. • Served models load as soon as you turn on Serve, so the first reply starts in under a second. • Models you serve stay ready for your apps. If chat, a Run window or a flow needs that memory, MLXUI asks before unloading one. • Smarter memory handling: every model is counted at its real size, and a model is never unloaded while it’s working, which prevents slowdowns and swapping. • A status card on Home shows what’s being served at a glance. • As before, the Local Server listens only on this Mac, and nothing leaves your device.

1. Qwen3.8-27B, a larger local chat model This 27-billion-parameter model runs fully on your Mac. It has a context of up to 256K tokens. Use the 3-bit build on Macs with 18–24GB of memory. Use the 4-bit build on Macs with 32GB or more. Find it in Browse → Chat & Text. 2. Chat with your files - Attach folders to a chat. The chat tools can use only those folders. - Add files from the message box, or drag them in. - A new sidebar shows the files the model read, created, or edited. Click a file to preview it, open it, or show it in Finder. - Edits show as a red and green diff. You see each change before you continue. - The app suggests a folder when a chat needs one. Click once to attach it. 3. Edit files with your AI assistant Ask for a text edit and check the change before the app saves it. This feature is off by default. Turn it on in Settings. 4. Export your chats Save a chat as Markdown or JSON, or copy it to the clipboard. Use the export button in the chat header or in the sidebar. 5. Local Server: ready for coding agents - Tool calling: Apps such as OpenCode and Continue can use your local models. Your app runs the tool calls. MLXUI does not. - Agent-ready badge: We tested Qwen3-8B, Qwen3.5 9B, and Ministral 3B with a coding agent. - Thinking on request: Qwen3 and Qwen3.5 models show their reasoning only when your app asks for it. - Clearer errors: You get a plain error before the reply starts if a prompt is too long or a model's template refuses it. - Easier setup: The Connect sheet has setups for OpenCode and Continue. 6. Faster repeat prompts Coding agents send the same long instructions with each message. MLXUI now remembers that work. In our tests, a repeated 8,000-token prompt starts to answer in under 1 second. Before, it took 30 seconds or more. The cache stays in memory only. The app never writes it to disk and frees it when a model needs the space. 7. Fixes - Tool calls are more reliable with Llama 3.1. - The window no longer opens too small on first launch.

What’s New Use Local AI models in other apps * Serve any downloaded chat model or Apple’s on-device model. * The server runs only on your Mac. It does not send your data to the cloud or save your prompts. * Models are loaded once and shared by Chat, Workflows, and Server. CLM 8B decision model * CLM 8B is now available for on-device decisions. * It can classify, choose, or score without generating text. * Use it in Classify, Gate, and Score workflow steps. Note: CLM 8B needs about 12 GB of memory. Settings * Use your own API keys for cloud models and search providers. * Providers and Privacy show what data can leave your Mac.

Laya, an on-device decision model. Give Laya a block of text and a question — a set of choices, a rating scale, or yes/no — and it returns an answer with a confidence score. Useful for fast routing and triage tasks like ticket classification. Runs entirely on your Mac; nothing leaves your device.

What's New 1. Added Qwen3.5-9B, a new local chat model with up to 262K tokens of context — great for long documents and extended conversations, fully on-device. 2. Two new automations in AI Workflows → Basic Gallery: Web Page Summary (paste a URL, get a one-paragraph summary) and Extract Web Page Links (pulls the top stories off a news page into a clean HTML digest).

New: AI Image Upscaling with SeedVR2 Turn low-resolution images into sharp, high-resolution ones — right on your Mac. SeedVR2 3B is a one-step super-resolution model that runs entirely on-device via MLX, with no upload, no account, and no internet connection required. It needs only ~4.5 GB of RAM, so it runs well even on lower-memory Apple Silicon Macs. Also in this release: minor stability improvements.

1. GLM-OCR This update adds GLM-OCR, a new on-device vision model for turning images of documents into structured text. Point it at a screenshot, scan, or photo of a spreadsheet, invoice, or report and it extracts the tables as clean structured data you can copy out — rows and columns preserved, not just a flat wall of text. Like every model in the app, GLM-OCR downloads once and runs entirely locally on your Apple Silicon Mac through MLX: your images never leave your device, and there's no account, subscription, or network round-trip. You can run it directly from the model's Run screen or drop it into an AI Workflow to batch-process a folder of documents. 2. Table From Image - AI Workflows Drop in a screenshot of a table — an invoice, a budget, a report page — and GLM-OCR/PaddleOCR-VL reads it in its table mode, returning rows and columns as structured markup instead of a wall of run-together text.

mlx-workflow - A Readable AI Workflow you can Run locally to automate Local LLM in an app. Readable AI Workflows supported -Transcribe Audio -Summary from Audio Click on the sidebar Automate->AI Workflows->Basic Gallery to try out.

SAM3 (4bit-mlx) Integration: Use Meta's powerful interactive image segmentation.

Type a musical description and tap Generate — MusicGen-small produces a WAV file directly on your Mac using Apple Silicon, with no internet connection required after install. Choose Short (~5 s), Medium (~10 s), or Long (~20 s) output; play back inline or Save WAV to keep the file. The model installs in one click from the catalog (~2.5 GB including the bundled audio codec). Non-commercial use only (CC-BY-NC-4.0).

SDXL-Turbo image generation A second on-device image-generation model joins FLUX.1 — Stability AI's SDXL-Turbo turns a text prompt into a 512×512 image in seconds, right on your Mac. No Python, no cloud, no API keys. - Fast 4-step generation with live progress - Save your generated image as a PNG

FLUX.1 image generation is here. Install FLUX.1-Lite-8B from the catalog and generate images from a text prompt — entirely on your Mac, no internet connection required after download, no API fees, no Python. The full pipeline (text encoding, diffusion, image decode) runs natively on Apple Silicon via MLX. This release also fixes a bug where canceling a model install mid-download could leave partial files on disk, wasting storage space.

Chat models can now use tools: browse the web, run calculations, transcribe audio, read images, and work in folders you grant — all locally, with your approval. New LLM Models -Ternary-Bonsai-27B