Skip to main content
PromptQuorumBuilt for humans. Structured for AI.
Home/Power Local LLM/MLXHub Review: On-Device MLX Chat and LAN Server for iPhone and iPad
Mobile & Edge LLMs

MLXHub Review: On-Device MLX Chat and LAN Server for iPhone and iPad

·8 min read·By Hans Kuepper · Founder of PromptQuorum · Discovery engine for open-weight & open-source AI
Read in:
🇺🇸en🇩🇪de🇫🇷fr🇯🇵ja🇨🇳zh🇪🇸es🇧🇷pt🇸🇦ar🇰🇷ko

MLXHub is a freemium, closed-source iOS and iPadOS app that runs open language and vision models offline on Apple silicon, can expose them as an OpenAI-compatible server on your Wi-Fi, and can pool several of your devices to load a model none of them fits alone. Per its App Store listing it needs iOS or iPadOS 26.0 or later and is free to download, with MLXHub Plus as in-app purchases. This review covers version 2.2.0 and is based on the public listing and official site, not hands-on testing.

MLXHub is an iPhone and iPad app that runs open models on the device with Apple's MLX framework, can serve them to other apps on your network, and can split one large model across several of your own devices. This review covers what its App Store listing and official site document, what is unclear, and how it compares with other iOS local-AI apps.

MLXHub Review: On-Device MLX Chat and LAN Server for iPhone and iPad

Key Takeaways

  • Version reviewed: 2.2.0, as shown on the App Store listing.
  • Price: free to download; MLXHub Plus is listed at $6.99 per month, $44.99 per year and $229.99 lifetime.
  • Platform: iOS and iPadOS 26.0 or later; the official site asks for Apple silicon and 6 GB of RAM or more.
  • Standout features: LAN server, multi-device model splitting, tools and skills, Hugging Face mirrors.
  • Openness: closed source; no public repository was found.

📍 In One Sentence

MLXHub, by developer Juan Colilla, is a freemium iPhone and iPad app that runs open models offline on Apple's MLX framework, serves them to your network, and can split a model across several of your devices.

💬 In Plain Terms

Think of it as a model runner that lives in your pocket: chat works with no internet once a model is downloaded, and the same device can act as a small private AI server for other apps at home.

Get MLXHub

MLXHub is distributed through the Apple App Store. The download is 73 MB and the app is a free download with in-app purchases.

This review is a companion to PromptQuorum's Local LLM Software Directory, which lists MLXHub alongside other mobile and desktop local AI tools.

Channel
Get It
Apple App StoreMLXHub: Local AI & LLM Server on the App Store
Official sitemlxhub.app
Privacy policyMLXHub privacy policy

No public source repository was found. There is no Android build; the listing covers Apple platforms only.

MLXHub at a Glance

Attribute
MLXHub
PlatformiOS, iPadOS (26.0+)
PriceFree, in-app purchases
LicenseClosed source
Runs fully offlineYes, after a model is downloaded
In-app model downloadsYes, from Hugging Face
Image inputYes, vision-language models
Local API serverYes, OpenAI-compatible on your LAN
Voice input / outputNot stated in the listing

Attributes follow the mobile-chat comparison used in the Local LLM Software Directory. They come from the App Store listing and official site, which do not cover every feature.

What MLXHub Is

MLXHub is an on-device model runner first and a chat client second. The listing pitches "private offline chat & tools" and says it lets you "run supported larger models and serve chat and embeddings across your local network". The official site says it is built natively in Swift on mlx-swift, Apple's MLX framework for Apple silicon.

It is published by Juan Colilla, with the copyright line "DreamFoundries EU". Do not confuse it with mlxhub.ai, a separate macOS-only app by a different developer that shares the name.

This review draws on the App Store listing and the official site only. It does not include hands-on testing, so speed, stability and answer quality are not rated here.

How to Get Started

Setup is an App Store install, then a model download from Hugging Face. The listing does not document the first-run flow, so the steps below follow what the listing and site state.

  1. 1
    Check device and OS
    Why it matters: The listing requires iOS or iPadOS 26.0 or later, and the official site asks for an Apple silicon device with 6 GB of RAM or more.
  2. 2
    Install MLXHub
    Why it matters: Get it from the [App Store](https://apps.apple.com/us/app/mlxhub-local-ai-llm-server/id6766485144); the download is 73 MB.
  3. 3
    Download a model
    Why it matters: Pick an open model from the in-app Hugging Face catalog. The app indicates whether it fits, but Hugging Face metadata is sometimes inconsistent.
  4. 4
    Chat, then optionally turn on the LAN server
    Why it matters: Once a model is downloaded, chat works with no connection. The LAN server is optional and lets other apps on your Wi-Fi use the loaded models.

Features

MLXHub's selling point is breadth around the model runner. Everything below comes from the App Store listing and the official site.

  • Models. Downloads open text and vision-language models from Hugging Face; the listing names Qwen3.6 35B-A3B, Bonsai 2 27B and Gemma models. Text-only chats load vision components only when needed, to save memory.
  • Memory and skills. A memory system keeps facts from conversations, and you can create reusable custom skills.
  • Native tools. Predefined tools reach Apple frameworks such as Calendar, Reminders, Health, Photos and the music library; the listing gives no detail on what each tool can do.
  • Speed and fit. KV caching speeds up multi-turn chats, and a RAM stress test benchmarks your device's memory headroom before you load a model.
  • Model switching and mirrors. Switch between text and vision models mid-conversation, and use a community Hugging Face mirror when Hugging Face is blocked.
  • Large-model streaming. The listing mentions NAND expert streaming for larger models such as Qwen3.6 35B-A3B, without further detail.
  • Languages. The interface is available in English plus 11 other languages.

LAN Server and Distributed Inference

The LAN server turns your device into a small OpenAI-compatible endpoint. The listing names the /v1/chat/completions and /v1/embeddings routes. Per the official site it also serves a lightweight chat page that anyone on the same Wi-Fi can open in a browser to talk to the loaded models, and it binds to local interfaces only, with no port forwarding.

Distributed inference pools several of your own devices. One device hosts and the others join, each taking a slice of the model's layers, so a model that none of them fits alone can load. The site says the devices are paired and the traffic encrypted, and that nothing leaves the network. Throughput over Wi-Fi is not published, so expect it to be slower than a single device that fits the model.

Device Requirements

MLXHub requires iOS or iPadOS 26.0 or later, an Apple silicon device and 6 GB of RAM or more. The official site recommends an A17 Pro or M1 chip, or newer. The App Store listing also shows macOS 26.0 (Apple M1 or later) and visionOS 26.0 availability for the same app, without describing a separate desktop version.

RAM is the real limit. Apple enforces stricter app memory limits on iPhone than on iPad, so larger models fit more often on an iPad. The site warns that a model that does not fit can occasionally still be downloaded, and that the app tries to stop it loading before iOS shuts the app down, without a guarantee.

Pricing and Privacy

MLXHub is free to download, with MLXHub Plus listed at $6.99 per month, $44.99 per year and a $229.99 lifetime purchase. Neither the listing nor the site says which features Plus unlocks, so check the paywall in the app before relying on a feature. Models are downloaded free from Hugging Face, with no per-token fees.

The official site says inference is fully on-device, there are no accounts and nothing identifies you. It also states that since version 2.0.0 the app sends anonymous analytics and diagnostics, covering which features you use, device class and whether a model loaded, and that you can switch them off in Settings.

The app is closed source and this review did not inspect network traffic, so the privacy claims are the developer's own. If you enable the LAN server, the data on it is as private as your Wi-Fi network.

Trade-Offs: Benefits vs. Limitations

Offline MLX inference

What it means in real use:
Chat with open models with no connection, using Apple's own framework.
Limitation / caveat:
Needs iOS 26 and 6 GB of RAM; iPhone memory limits cap model size.

LAN server

What it means in real use:
Other apps and browsers on your Wi-Fi can use the loaded models.
Limitation / caveat:
The device must stay awake and on the network; no throughput figures are published.

Multi-device splitting

What it means in real use:
Load a model that no single device of yours can hold.
Limitation / caveat:
Speed over Wi-Fi is undocumented and needs several compatible devices.

Tools, memory and skills

What it means in real use:
Reach Calendar, Reminders and Health from a chat.
Limitation / caveat:
The listing gives no detail on what each tool does.

Free to start

What it means in real use:
No per-token fees; models are free from Hugging Face.
Limitation / caveat:
Plus costs up to $229.99 lifetime, and what it unlocks is not stated.

MLXHub vs. Alternatives

MLXHub

Platforms:
iOS, iPadOS
Price and license:
Freemium, closed source
Model source:
Hugging Face (MLX)
Key difference:
LAN server and multi-device model splitting; not open source

Locally AI

Platforms:
iOS, Mac
Price and license:
Freemium, closed source
Model source:
In-app model catalog
Key difference:
Simpler on-device chat; no LAN server is listed
Articles about Locally AI (4)

Also mentioned in:

Private Mind

Platforms:
iOS, Android
Price and license:
Free, MIT
Model source:
Hugging Face downloads
Key difference:
Open source and on Android too; no LAN server is listed
Articles about Private Mind (1)

Also mentioned in:

Oscilla

Platforms:
iOS
Price and license:
Free, closed source
Model source:
40+ in-app models
Key difference:
Larger fixed catalog; runtime not named
Articles about Oscilla (1)

PocketPal AI

Platforms:
iOS, Android
Price and license:
Free, MIT
Model source:
GGUF models
Key difference:
Open source, llama.cpp-based, cross-platform
Articles about PocketPal AI (10)

+20 more not shown

Platform, price, and feature details for third-party apps change frequently. Verify current specifics on each app's own listing before deciding.

Compare in directory

Who Should Use MLXHub

  • iPad and recent iPhone owners who want the largest models their device can hold. MLX, streaming and device pooling all target that limit.
  • Tinkerers who want a pocket AI server. The OpenAI-compatible LAN endpoint lets scripts and other apps on your network use the device.
  • People with several Apple devices. Multi-device splitting is the feature that most alternatives do not list.

Who Should Not Use MLXHub

  • Anyone who needs open source. The code is not public; Private Mind and PocketPal AI are open source.
  • Android users. Only Apple platforms are listed; PocketPal AI and Private Mind both run on Android.
  • Devices below iOS 26 or under 6 GB of RAM. The listing and site set those floors.
  • Anyone who wants a stated feature list for the paid tier. What MLXHub Plus unlocks is not documented; try the free tier first.

Frequently Asked Questions

Is MLXHub free?

It is free to download with in-app purchases: MLXHub Plus at $6.99 per month, $44.99 per year, or $229.99 lifetime, as shown on the App Store listing.

Who makes MLXHub?

Juan Colilla, with the copyright line "DreamFoundries EU". The mlxhub.ai app is a different, macOS-only product.

Does MLXHub send my chats anywhere?

The site says inference is fully on-device. Anonymous usage and diagnostics data is sent since version 2.0.0 and can be switched off in Settings.

Does MLXHub work on Android?

No. The listing covers iPhone, iPad and, per the store, Apple M1 Macs and Apple Vision Pro compatibility only.

Which models can I run?

Open MLX-format text and vision models from Hugging Face. The listing names Qwen3.6 35B-A3B, Bonsai 2 27B and Gemma models.

Verdict

MLXHub is a good fit on paper for someone with a recent iPad or iPhone who wants offline MLX chat plus the option to use the device as a local endpoint or to pool several devices for a bigger model.

The caveats are transparency and age: it is closed source, the Plus features are undocumented, speed over multi-device splitting is unpublished, and the listing showed only seven ratings at the time of review.

If you want open source, start with Private Mind. If you want the LAN server or multi-device splitting, try MLXHub's free tier first.

Sources

← Back to Power Local LLM