Qwen3.6 35B-A3B, a model that until now would not fit on a phone, runs on your iPhone and your iPad. It gets there through NAND expert streaming: instead of holding the whole model in memory, it fetches each expert the moment that expert has something to say, and lets go of the ones that don't. Supported vision models also stop carrying weight you are not using. While you type, the vision side stays asleep and only wakes up when you attach a photo, so a text conversation takes noticeably less memory than it used to. The assistant can reach into your apps: your calendar, your reminders, your Health data, your photos or your music library, always with your permission, call by call. And whatever you explain once can be saved as a Skill, so you never have to explain it again. It has memory now, too. What is worth knowing about you carries over to your next conversations, and you decide what stays: review, pin or delete any memory from Settings. Distributed inference across several of your devices on the same local network is no longer an experiment. It is simply part of the app. Joining a mesh is still free. And your iPhone or iPad can serve the rest of your home: chat and embeddings for the apps and scripts on your local network, over the standard routes /v1/chat/completions and /v1/embeddings, with authentication and local discovery. Support depends on the model you pick and the device you have. Write tools, Skills, letting the assistant save memories on its own, hosting a mesh and the LAN server are part of MLXHub Plus, which is exactly what keeps all the rest coming. Thank you for being here.

Qwen3.6 35B-A3B, a model that until now would not fit on a phone, runs on your iPhone and your iPad. It gets there through NAND expert streaming: instead of holding the whole model in memory, it fetches each expert the moment that expert has something to say, and lets go of the ones that don't. Supported vision models also stop carrying weight you are not using. While you type, the vision side stays asleep and only wakes up when you attach a photo, so a text conversation takes noticeably less memory than it used to. The assistant can reach into your apps: your calendar, your reminders, your Health data, your photos or your music library, always with your permission, call by call. And whatever you explain once can be saved as a Skill, so you never have to explain it again. It has memory now, too. What is worth knowing about you carries over to your next conversations, and you decide what stays: review, pin or delete any memory from Settings. Distributed inference across several of your devices on the same local network is no longer an experiment. It is simply part of the app. Joining a mesh is still free. And your iPhone or iPad can serve the rest of your home: chat and embeddings for the apps and scripts on your local network, over the standard routes /v1/chat/completions and /v1/embeddings, with authentication and local discovery. Support depends on the model you pick and the device you have. Write tools, Skills, letting the assistant save memories on its own, hosting a mesh and the LAN server are part of MLXHub Plus, which is exactly what keeps all the rest coming. Thank you for being here.

Your personal AI platform reaches version 2.0.0 just two months after launch! We have made a huge number of changes so you can enjoy the app more, without having to make your life complicated trying to understand it. To begin with, the interface has been polished to fit more information into the same space. We have represented concepts visually to reduce cognitive load and improved other areas that needed it. The app now comes to life not only through its orbs —rendered entirely with Metal— but also through more refined animations. An experimental feature has been added: distributed inference! You can now run larger models that do not fit on your device. Have an iPad sitting in a drawer? Add it to the mesh. Your partner’s iPhone? Add that too, and connect devices like puzzle pieces to build an increasingly capable inference server. We admit the previous value proposition was not clear: the MLXHub Plus subscription offered very little, but that changes now. MLXHub Plus now includes LAN Server, distributed-inference hosting, unlimited models, app customization, and future features that will continue adding value. Do not want to subscribe? No problem: there is also a one-time purchase! Version 2.0.0 includes many under-the-hood changes that you will not easily notice, such as fixes for most VLMs that sometimes caused issues, automatic cleanup of partial files generated by the app, improvements to model downloads, more detailed model pages, and many framework improvements that make MLXHub the best app for carrying cutting-edge AI in your pocket: without internet, without model removals, and without the fear that some orange-haired character may feel like blocking your access from an armchair with paper and pen. Enjoy these new features and the immense number of changes still to come. Do not hesitate to contact us with any question or suggestion. Best regards, Juan

Fixed a bug that prevented setting whether a model should reason or not via the LAN server.

MLXHub now has a more refined visual experience throughout the app, with new activity orbs, improved contrast, and theme color options for Plus. We also polished animations and transitions so chat, models, and the LAN server feel smoother. This update improves performance and responsiveness during generation, scrolling, and server use.

Open MLX models directly from Hugging Face. Tap MLXHub on a compatible Hugging Face model page to open its details in the app and download it. If MLXHub isn’t installed, you’ll be taken to the App Store.

Fixed the LAN server port field: you can now type any port number without it jumping to another value on its own. Also fixed the keyboard getting stuck open with no way to dismiss it on several settings screens — swipe the list or tap Done to close it.

MLXHub expands compatibility with 1-bit quantized MLX models, including Bonsai 27B. This update integrates new versions of mlx-swift and mlx-swift-lm to load and run these models directly on your device.