NewiPhone · iPad · iPhone Duo · Mac

MLX Workbench

A private AI chat that runs on your iPhone, iPad and Mac, with open models and no cloud AI.

Free download · Chat with PRO, a one-time purchase · No account · Requires iOS 18.6 or a Mac with Apple silicon

A four-page PDF quote attached to the chat: its key points and the hidden cost on page 3
The chat on iPhone 17 Pro, on the device and offline: the monthly payment of a mortgage, worked out with the formula
A photo of eggs, spinach, feta, tomatoes and an onion, and the recipe suggested by a model that reads images

MLX Workbench is a private AI chat that runs on your own devices, with open models and Apple’s open-source MLX framework. Your conversations stay yours and never go through a server: no cloud AI, no account, no subscription.

Attach a PDF, ask about a photo, paste a link or dictate; keep a few notes about you that every chat starts from; and when a model is too big for your iPhone, use it on your Mac, at home and away. The free benchmark is still here: every device runs the same standard test, so you can see where yours stands.

What's new · 1.2Private AI chat, your Mac’s models on iPhone and iPad, documents, photos, links and memory

What it does.

A private chat

Open models such as Qwen, Llama, Gemma, Mistral and Phi, or Apple’s on-device model. Once a model is on your device, it works offline.

Your Mac’s models

Use your Mac’s models from iPhone and iPad, even those too big for your iPhone, at home and away. Pair once with a QR code; the chat travels encrypted, straight between your devices.

Documents

Attach PDFs and text files and ask about them. Scanned pages are recognized on the device.

Photos

Ask about a photo, the camera or an image in the clipboard, with a model that reads images.

Links and voice

Paste a link and the page’s article is read into the chat. Dictate instead of typing, with on-device speech recognition.

Memory

Notes about you that every chat starts from. You review them, edit them and decide what to keep.

Thinking and control

See the reasoning, then the answer, with clean math formulas. Edit, regenerate, compare versions, stop and continue; summarize long chats to free up memory.

Search and export

Find a chat by its words. Export a chat as Markdown, JSON or a ChatGPT-style archive.

Models that fit

What is ready, what is downloading and what fits your device’s memory, with the speed to expect. Downloads continue in the background, with pause and resume.

Stats for every reply

Time to first token, generation and prefill speed, cache, peak memory and thermal state, live while the reply is written.

Benchmark and Community, free

The same standard test on every device: compare yours with everyone else’s, model by model, including Apple’s model and the speed measured in real chats.

Your privacy

No account, no ads, no tracking. Your conversations stay on your devices and never go through a server.

On iPhone, more than a chat.

Have a web page explained in plain words, keep a few notes about you, see what every reply took, and pick the models that fit your iPhone.

Chat

A Wikipedia page on neural networks attached to the chat and explained as if to a ten-year-old
Links
What I remember about you: four notes every chat starts from, with the switch to use memory in chats
Memory
Reply details: time to first token, generation and prompt reading speed, peak memory, the speed chart and the cache
Reply stats

Models and benchmark

The Models tab: models ready to use, one downloading and those recommended for this iPhone, with size and expected speed
Models
The standard benchmark on iPhone 17 Pro, next to the community median, with a button to try the model in chat
Benchmark

Your Mac’s models, on iPhone and iPad.

On the Mac, MLX Workbench is the same chat in a window, with your chats in the sidebar. Turn on “Host chats on this Mac” and scan its QR code once: your iPhone and iPad can use the Mac’s models, even the biggest ones, at home and away, and the chat travels encrypted, straight between your devices.

From iPhone

The model list on iPhone: Apple’s model, Qwen3 32B and 8B on the Mac Studio over the same Wi-Fi, and the models on the iPhone
Choose a model
A question about a positive medical test answered by Qwen3 32B running on the Mac Studio
A model on the Mac

On the Mac

MLX Workbench on the Mac: the chats in the sidebar and a mortgage question answered with the formula
The chat on the Mac
Settings on the Mac: Host for iPhone and iPad, reachable from outside the home, replying to an iPhone, two connected devices
Host for iPhone and iPad

On iPhone Duo, closed and open.

Closed, the chat you know, easy to use with one hand. Open, your chats sit next to the conversation, like in Mail.

Closed

iPhone Duo closed: a chat about a mortgage
Chat
iPhone Duo closed: a recipe from a photo of ingredients
Photos
iPhone Duo closed: the notes the app remembers about you
Memory

Open

iPhone Duo open: the list of chats next to the conversation
Chats and conversation
iPhone Duo open: the sidebar next to the Models tab, with a download in progress
Models

On iPad, your chats in the sidebar.

Like in Mail, the sidebar keeps every chat one tap away, and Return sends your message with a hardware keyboard.

iPad in landscape: the chats in the sidebar and a mortgage question answered with the formula
Chat
iPad in landscape: the Models tab and a model’s page, with how much memory it needs and its expected speed on this iPad
Models

Frequently asked questions

Is it free?

Downloading it is, and so are Benchmark, Community and the Models tab. The chat is part of MLX Workbench PRO: a one-time purchase, no subscription, shareable with Family Sharing. 30 free replies let you try it first, shared by your devices and the connection to your Mac.

Does one purchase cover all my devices?

Yes. PRO unlocks the chat on iPhone, iPad and Mac with the same Apple Account, and Family Sharing extends it to your family.

Do my conversations go through a server?

No. Models run on your devices and your conversations stay there. With a connected Mac the chat goes straight to your Mac, encrypted; iCloud only keeps your Mac’s address and your free-reply count, never your conversations. When you paste a link, the page is downloaded from its website.

Does it work offline?

Yes, once a model is on your device. Downloading models and reading links need the internet, and a connected Mac has to be reachable.

Which models can I use?

Open models from the MLX community on Hugging Face, such as Qwen, Llama, Gemma, Mistral and Phi, and Apple’s on-device model where Apple Intelligence is available. The Models tab shows which ones fit your device’s memory and how fast they run on it.

How does the connected Mac work?

Open MLX Workbench on a Mac with Apple silicon, go to Settings › Host for iPhone and iPad, tap “Connect an iPhone” and scan the QR code with your iPhone or iPad. The Mac’s models then appear in the model list. At home they connect over Wi-Fi. Away from home, your home connection needs a public IP address and a router with UPnP or NAT-PMP turned on; the Mac checks this and tells you. Behind carrier-grade NAT (CGNAT) it works on the same Wi-Fi only.

What does the standard benchmark measure?

Prefill speed on a fixed 512-token prompt, then generation speed over exactly 128 tokens with greedy decoding, plus time to first token, per-token latency and memory. There is one warm-up run and three measured runs, and the medians are reported. Because the input and settings never change, results from different devices can be compared.

What is shared with the community?

Only while “Share results with the community” is on: standard benchmark results and, once a week, the median chat speed of each model, with the device model and its hardware, the model, and the app and OS version. Numbers only: never the text of your conversations, and nothing that identifies you. You can turn sharing off in the Benchmark tab or in Settings.

Why are my results slower than the community’s?

Low Power Mode, a hot device and other apps at work all slow generation down. Runs in Low Power Mode or ending on a hot device are left out of the community medians: let the device cool down and try again.

How much space do the models need?

The smallest, Qwen3 0.6B in 4-bit, takes about 350 MB; an 8-billion-parameter model in 4-bit takes about 4.6 GB. Apple’s on-device model is part of the system and needs no download.

Which devices does it run on?

iPhone and iPad with iOS 18.6 or later, and Macs with Apple silicon. On iPhone Duo the app follows the device, closed and open. Apple’s on-device model needs iOS 26 and a device with Apple Intelligence turned on.

MLX Workbench is made by Sofapps.

Sofapps is a small studio on the slopes of Mount Etna. We design, build and look after our own apps, for the people who use them every day.

All our apps

Need help?

Write to us: we read every message and reply personally.

Free download · Chat with PRO, a one-time purchase · No account · Requires iOS 18.6 or a Mac with Apple silicon