Release notes

dMLX v0.1.0-beta3

A guided first start, six more image models, and a new look.

Highlights

First-run setup. On first start dMLX recommends an image model and an agent model for your Mac’s memory, shows the total download, and checks that there is enough free disk space. You can change either model, or skip the agent. While the models download you can continue in the background; you can start as soon as the image model is ready. The setup shows each model’s size before it is downloaded, and a message sent before the image model is ready stays in the box, with the download’s progress.
More image models. Settings → Models → Add Model now offers ERNIE Image Turbo, MOSS Image, FLUX.3 Klein, Boogu Image Turbo, Iris 3B Turbo and Krea 2 Turbo. Every family runs through one adapter and gives the same images as mflux’s own pipelines. Some of these take minutes per image.
New look. The Still Water logo and app icon, blue colors, and Settings → General → Appearance: System, Light or Dark.
Crop / Draw. Erasing paint on the canvas now matches the exported image. The agent knows what a prepared (marked) image is, and each edit is labeled with the image it was made from.
Chat. A chat opens at the latest message, and it scrolls to a message being planned and to the agent’s question.
Memory. Sizes under 1 MP are planned as 1 MP, so small images no longer reserve too little memory. The text encoder is freed when encoding ends (lower peaks for FLUX.3 and MOSS).
mflux 0.24.

Requirements

Apple Silicon, macOS 15+, 16 GB memory or more. About 7–11 GB of disk for the image model and the agent.

Install

Open the .dmg and drag dMLX onto the Applications folder (a .zip of the same app is also attached), then right-click → Open the first time (beta builds are not notarized). The first-run setup downloads the models.

Known issues

See Known issues.

dMLX v0.1.0-beta2

Adds a planning agent: a small vision-language model on your Mac reads every message, looks at the images, and decides what to do.

Highlights

Planning agent. Decides whether a message asks for a new image or an edit, and which images to edit or use as references (“make image 1 black and white”, “put the teapot from the first one in the boat”). It can look at older images in the chat and read further back when it needs to. For an edit it can write a detailed description of the result for the image model.
It talks in the chat. A greeting or a question about the app gets a short reply; when a message is unclear the agent asks one question. Answer it, or just ask for something else.
Auto picks the image model. In Auto the agent chooses among your installed models by what the message needs and the free memory; a model you pick yourself is never overridden.
Choose the agent model. Settings → Intelligence → Agent: Qwen3-VL 8B (default from 24 GB), Qwen3.5 9B, Ministral 3 3B (default under 24 GB) or Qwen3-VL 4B; 2.8–6.0 GB downloads. It is loaded only while it plans and freed before the image job starts.
Messages are never dropped. A sent message shows in the chat with “Planning…”; you can keep typing. A message sent from another chat waits for the running image, then is planned.
Qwen Image 2.1 (experimental, research license). Can be added in Settings → Models: generate and edit with up to 10 ordered references, best at exact words in the image. It needs about 9.5 GB free and takes about 4 minutes per image; in Auto only the agent picks it.
Removed: the per-image Attach button. Refer to images by number instead, or attach files from your Mac.

Requirements

Apple Silicon, macOS 15+, 16 GB memory or more. Plus 2.8–6.0 GB disk for an agent model.

Install

Open the .dmg and drag dMLX onto the Applications folder (a .zip of the same app is also attached), then right-click → Open the first time (beta builds are not notarized). Download an agent model in Settings → Intelligence → Agent; without one, messages are routed by rules.

Known issues

See Known issues.

dMLX v0.1.0-beta1

First public beta: chat to create and edit images, fully on your Mac.

Highlights

Chat-style creation. Describe an image; follow up in plain words to edit the latest result. The composer shows what will happen (“Will edit image 2 · FLUX.2 Klein 4B”) before you send.
Model menu in the message box. Auto by default, or pick a model; the line above the box says which model the next message uses, and notes when it has to load a different one.
Add and remove models. Any FLUX.2 Klein or Z-Image model in mflux format, from a list of community versions or a pasted Hugging Face link; checked (family, size, license) before download.
Versions. Variations stack as v1, v2, v3 on the same turn; edit any version; use any result as a reference for the next edit.
Never hangs your Mac. Memory is checked before every job; model parts load one at a time when memory is tight, and when it is very tight the model streams from disk block by block (a new 1024×1024 image in about 1.5 GB). Below 32 GB the decoder runs in exact low-memory bands; jobs stop automatically if macOS memory pressure turns critical or apps start being read back from swap.
Warm follow-ups. Models stay loaded between turns and prompt embeddings are cached; a follow-up edit does zero model loads.
Recovers by itself. If the engine crashes it restarts; the interrupted turn offers Retry.
Local OpenAI-compatible API for other apps and agents (/v1/images/generations, /v1/images/edits).

Models

FLUX.2 Klein 4B (default, generate + edit, Apache-2.0) and Z-Image Turbo (optional, Apache-2.0), 4-bit packs from mflux-community.

Requirements

Apple Silicon, macOS 15+, 16 GB memory or more.

Install

Unzip, move to Applications, then right-click → Open the first time (beta builds are not notarized).

Known issues

See Known issues. Most important: the 16 GB target still needs validation on a real 16 GB Mac — reports welcome.