Quick Facts
- Apps compared: PocketPal AI, Private LLM, Locally AI (by LM Studio), Google AI Edge Gallery, MLC Chat, Apple Intelligence (system) — plus LLM Farm, currently not on the App Store (README: "temporarily unavailable").
- Last upstream updates (checked 2026-10-23): PocketPal AI v1.20.1 (19 Oct 2026) · Locally AI v1.66.0 (15 Oct 2026) · Google AI Edge Gallery v1.0.13 (8 Oct 2026) · Private LLM v1.9.16 (14 Sep 2026) · MLC Chat iPhone app v1.6 (27 Sep 2024 — stale) · LLM Farm release 1.4.3 (27 Jan 2025), last code push 30 Jan 2026 — stale.
- Reference devices for speed figures: iPhone 16 Pro (A18 Pro, 8 GB RAM) and iPhone 17 Pro (A19 Pro, 12 GB RAM). Speed figures are rough orientation ranges, not measurements by PromptQuorum. The iPhone 18 Pro, 18 Pro Max, and iPhone Duo (A20 Pro) have no speed figures here.
- Inference engines: llama.cpp (PocketPal AI, LLM Farm), Apple MLX (Locally AI), MLC LLM with Metal (MLC Chat), Google LiteRT-LM (Google AI Edge Gallery), proprietary on-device runtime (Private LLM, Apple Intelligence).
- Minimum iPhone for 3B+ models: iPhone 14 Pro (A16, 6 GB RAM) for 1.7B; any 8 GB+ iPhone (15 Pro, 16 series, 16e, 17 series, 18 Pro) for 3B–4B.
- Recommended free app: PocketPal AI — App Store, open-source, model flexibility.
- Paid option: Private LLM — $4.99 one-time (iPhone + iPad + Mac), Shortcuts + Siri support.
- Offline: After the model is downloaded, on-device chat in all six works without WiFi; features that fetch live data (for example Google AI Edge Gallery's Wikipedia and maps skills, or Locally AI's optional LM Link) need a connection. Per the README or listing of PocketPal AI, Private LLM, Locally AI and Google AI Edge Gallery, on-device chat makes no cloud calls (not verified for MLC Chat or LLM Farm); Apple Intelligence can use Private Cloud Compute.
Which iPhone App Should You Install First?
For most users: PocketPal AI from the App Store, then download Phi-4 Mini (3.8B Q4_K_M) — or a newer small model such as Gemma 4 E2B/E4B or Qwen3.5 4B. This combination works on iPhone 15 Pro and newer (8 GB+ RAM); on a 6 GB iPhone such as the iPhone 14 Pro, download a 1.7B model such as Qwen3 1.7B instead (see Models by iPhone Tier). It costs nothing for chat, and produces usable output for everyday chat, summarisation, and quick drafting. Choose a different app only when you have a specific need it does not cover.
📍 In One Sentence
For most iPhone users in 2026, install PocketPal AI (free, App Store) and download Phi-4 Mini or a newer small model such as Gemma 4 E2B — it covers everyday chat, summarisation, and drafting — with a 3B–4B model on 8 GB+ iPhones and a 1.7B model on 6 GB iPhones.
💬 In Plain Terms
Six iPhone options run AI on your phone in 2026; the chat apps keep chat on the device once the model is downloaded, and Apple Intelligence can also use Apple's Private Cloud Compute for larger requests. PocketPal AI is the recommended free starting point — install it, download a 2.7 GB model file once, and you have a private chat assistant that works on the train without WiFi. Private LLM is the paid option if you want Siri to talk to the local model. Locally AI, from the LM Studio team, is the free modern alternative with Shortcuts support, and Google AI Edge Gallery is Google's free app for its Gemma 4 models. The iPhone build of MLC Chat is no longer updated. Apple Intelligence is built into iOS but not a chat app. PocketPal AI fits most use cases.
Decision: Which iPhone Local AI App?
Use a local LLM if:
- •You want a free chat AI that works offline → PocketPal AI
- •You want Siri to talk to your local model → Private LLM
- •You want a free MLX app with Shortcuts and Apple Foundation Models access → Locally AI
- •You want Google's own Gemma 4 with thinking mode and agent skills → Google AI Edge Gallery
- •You want to experiment with the MLC LLM toolchain and accept an iPhone build with no update since September 2024 → MLC Chat
- •You only need writing assistance in Mail / Messages / Notes → Apple Intelligence (built-in)
- •You want offline chat with document Q&A, an editable memory or a PIN-locked Kid Mode → MemoryAI
Use a cloud model if:
- •You need 70B+ model quality (Llama 3.3 70B, GPT-5.5 tier) → use cloud or remote-connect to a home machine
- •You need access to GPT-5.5, Claude Opus, or Gemini specifically → cloud apps (these are not available locally)
- •You need real-time vision or multimodal output beyond text → cloud (on-device multimodal is limited in 2026)
Quick decision:
- →Free + works for most users: PocketPal AI
- →Paid + iOS-native integration: Private LLM
- →Free + LM Studio ecosystem: Locally AI
- →Free + Google Gemma 4: Google AI Edge Gallery
- →Free + documents, memory, Kid Mode: MemoryAI
💡Tip: Install PocketPal AI first, even if you plan to pay for Private LLM later. Use PocketPal AI to test whether on-device inference on your iPhone tier feels fast enough for your use case. If yes, decide whether the Shortcuts and Siri integration in Private LLM is worth $4.99. If no, you saved $4.99 finding out before paying. PocketPal AI also has a built-in benchmark that reports tokens/sec and memory on your own iPhone — use it instead of trusting any table, including the one below.
iPhone App Comparison Table
The six options differ on three axes that matter to most users: cost, model flexibility, and integration with iOS. Speed differences exist but are smaller than the gap between any of these apps and a cloud LLM. Prices and versions were read from the US App Store listings and project pages on 2026-10-23.
📍 In One Sentence
PocketPal AI is the recommended free default, Private LLM is the paid option, Locally AI is the free MLX pick from the LM Studio team, Google AI Edge Gallery is Google's free Gemma 4 app, MLC Chat is dormant on iPhone, Apple Intelligence is system-integrated.
💬 In Plain Terms
The choice usually comes down to three things: do you want to pay for Siri integration (Private LLM), do you want the LM Studio ecosystem with Shortcuts for free (Locally AI), or do you want Google's own Gemma 4 with a thinking mode (Google AI Edge Gallery). For everything else, PocketPal AI is the default. The speed numbers in the table are indicative ranges for Q4_K_M quantisation — the standard for mobile inference in 2026 — and not measurements by PromptQuorum.
App | Cost | Tokens/sec (Phi-4 Mini, 16 Pro) | Privacy | Best for |
|---|---|---|---|---|
| PocketPal AI | Free (MIT) + optional in-app purchases | ~10–15 | On-device chat, no account (README) | Free default for most users |
| Private LLM | $4.99 one-time | No figure | On-device; no accounts or tracking (developer) | Shortcuts + Siri integration |
| Locally AI | Free (LM Studio team) | N/A — MLX catalog (no Phi-4) | On-device; no data collection (listing) | MLX + Shortcuts, LM Studio link |
| Google AI Edge Gallery | Free (Google) | N/A — LiteRT-LM, Gemma 4 | On-device after model download (listing) | Gemma 4, thinking mode, agent skills |
| MLC Chat | Free (open-source) | N/A — library predates Phi-4 | Local inference (project docs; unverified) | MLC LLM toolchain users (iOS app dormant) |
| LLM Farm | Free (build from source) | No figure | Local inference (project docs; unverified) | Developers (not on App Store, stale) |
| Apple Intelligence | Free (built into iOS) | N/A (system feature) | Local + Private Cloud Compute (Apple) | Mail, Messages, Notes assistance |
| MemoryAI | Free; no ads or in-app purchases | No figure | On-device after download; usage events and crash reports sent | Documents, editable memory, Kid Mode |
| fullmoon | Free (MIT) | No figure | On-device per maker; policy contradicts itself | Simple offline chat; no release since Jan 2025 |
| Cloak: Local AI | Paid once ($1.99, US store); no subscription | No figure | On-device after model download; optional survey and opt-in usage counts | Setup-free chat, Voice Memos, Recap |
Note on Apple Neural Engine (ANE) vs Metal vs MLX: PocketPal AI and LLM Farm use llama.cpp with Metal Performance Shaders for inference, which runs on the GPU. MLC Chat uses MLC LLM with deeper Metal optimisation, which the MLC LLM project describes as faster than llama.cpp-based apps on the same hardware (project claim; PromptQuorum has not measured this). Locally AI uses Apple MLX, Apple's own open-source array framework, which reaches similar GPU efficiency on Apple Silicon. Google AI Edge Gallery runs Gemma 4 through Google's LiteRT-LM runtime. Apple Intelligence uses ANE specifically for the system ~3B model, which is more power-efficient but less flexible than Metal-based inference. Since iOS 26, third-party apps can call that system model through the Foundation Models framework — Locally AI does exactly this — but custom GGUF/MLX models still cannot directly target the ANE. Apple states the A20 Pro in the iPhone 18 Pro has a Dual 16-core Neural Engine with two times the compute of before; that is an Apple claim, and PromptQuorum has not measured what it means for these apps.
💡Tip: Tokens/sec figures are rough orientation ranges for Q4_K_M models on recent iPhones, not measurements by PromptQuorum and not tied to an app version or test date. They assume Q4_K_M quantisation (the standard for mobile inference in 2026) and an idle iPhone with no other heavy apps running. Background apps can reduce throughput noticeably, and the iPhone 17 Pro typically runs faster than the iPhone 16 Pro. Measure your own phone with PocketPal AI's built-in benchmark.
PocketPal AI: Free Open-Source Default
PocketPal AI is the recommended starting point for most iPhone users in 2026. It is free to use, open-source (MIT; GitHub: a-ghorbani/pocketpal-ai), available on the App Store, and supports any GGUF model from Hugging Face. The app uses llama.cpp under the hood with Apple Silicon optimisations. Full review: PocketPal AI Review.
- What it is: an iOS/iPadOS and Android app that runs GGUF models locally using llama.cpp, published by Gooya UG. The README states prompts, responses, and documents are not uploaded to external servers; chat needs no account.
- Version and last update: v1.20.1, released 19 October 2026 (GitHub release); the App Store listing shows the same version. Recent releases: v1.19.0 (4 Oct), v1.20.0 (14 Oct), v1.20.1 (19 Oct 2026) — no stale warning.
- Install: App Store → "PocketPal AI". Free download with optional in-app purchases (premium Pals and "PocketPal Plus"). Requires iOS/iPadOS 15.1 or later; RAM, not the iOS version, sets the usable model size. Since v1.20.0 the app also offers an optional Sign in with Apple and account deletion.
- Adding a model: in the app, tap Models → "Add from Hugging Face" → search (e.g. "phi-4-mini-instruct-Q4_K_M") → tap to download. The model is stored in the app's local storage (~2.7 GB for Phi-4 Mini Q4).
- Generation speed (iPhone 16 Pro, indicative): Phi-4 Mini ~10–15 tok/sec, Llama 3.2 3B ~12–18 tok/sec, Gemma 3 4B ~7–10 tok/sec, Qwen3 1.7B ~18–24 tok/sec. These are rough orientation ranges, not measurements by PromptQuorum — use the built-in benchmark, which reports tokens/sec and memory on your own iPhone.
- Other features: Pals (custom assistants), on-device text-to-speech, Talents (calculator and date/time tools for capable models), and a Hugging Face model browser that supports gated models with an access token.
- Best for: users who want a free, App Store-installable chat app that needs no account and supports any community-released GGUF model.
💡Tip: The PocketPal AI model picker has a "Recommended" filter that, per the app's documentation, shows models that fit your device's RAM. Check it against your own phone — an oversized model can lead to iOS closing the app mid-response.
Private LLM: Paid Option with iOS Integration
Private LLM is the paid iPhone option in this guide ($4.99 one-time, no subscription). It is App Store only and ships with a curated library of optimised models. Its differentiator is iOS integration: the developer states it plugs directly into Siri and the Shortcuts app. Full review: Private LLM Review.
- What it is: a paid iOS app with a curated model library and deep iOS integration. Uses a proprietary on-device runtime optimised for Apple Silicon.
- Install: App Store → "Private LLM - Local AI Chat" (Numen Technologies Limited). One-time purchase $4.99 (no subscription; the listing shows no in-app purchases). Requires iOS/iPadOS 17.0 or later on an A12 Bionic chip or newer.
- Version and last update: v1.9.16, released 14 September 2026 (App Store version history); the previous release was 1.9.15 on 9 July 2026. Releases come every two to three months — under the three-month line, so no stale warning.
- Curated model library: privatellm.app advertises 140+ local models across iPhone, iPad, and Mac, naming families such as Llama, Qwen3, Gemma 3, and Phi 4; the subset that fits an iPhone is smaller. Less flexibility than PocketPal AI, but the models are pre-optimised for the app's runtime.
- Shortcuts: per the developer, Private LLM plugs into the Shortcuts app and can pass responses to other apps that support the x-callback-url specification. Useful for triggering local AI from a button on your home screen or via NFC tags.
- Siri integration: the developer lists Siri integration; a prompt routed to the on-device model works without an internet connection (internet is only needed for the first model download).
💡Tip: Per privatellm.app, one $4.99 purchase covers iPhone, iPad, and Mac, and the App Store listing says up to six family members can use it with Family Sharing. No subscription and no in-app purchases are listed — the listed price is the total cost. The developer states there are no accounts, tracking, or logs.
Locally AI: LM Studio's Free MLX App
Locally AI is a free chat app (its listing says it collects no data) developed by the LM Studio team (listed under Element Labs Inc), built on Apple's own MLX framework instead of llama.cpp. It is the only app here that exposes Apple's on-device foundation model as a chat interface. Full review: Locally AI Review.
- What it is: an iOS/iPadOS/macOS app using Apple MLX for inference, developed by the team behind the desktop LM Studio app.
- Install: App Store → "Locally AI". Free, no in-app purchases listed, no account required (per the listing), runs offline.
- Version and last update: v1.66.0, released 15 October 2026 (App Store version history); about one release a month in 2026 — no stale warning.
- Model library: Llama 3.2, Gemma 2/3/4, Qwen 3 and 3.5, DeepSeek, LFM 2.5, Bonsai, Ministral 3, and Apple Foundation Models — a broader and more current catalog than MLC Chat's.
- iOS Shortcuts: exposes a Shortcuts action, matching Private LLM's automation support at zero cost.
- LM Link: an optional feature that connects to LM Studio running on a Mac over what LM Studio describes as an end-to-end encrypted link, letting the iPhone app hand off to a larger model on a home machine when needed.
- Requirements: iOS/iPadOS 18.1 or later (macOS 26.0 or later for the Mac version). Apple Silicon-class efficiency benefits most on iPhone 15 Pro and newer.
💡Tip: Locally AI is the only app in this guide that can put Apple's own on-device foundation model behind a normal chat window — useful if you want to test what Apple Intelligence's model can do without hunting through Writing Tools menus. For third-party GGUF-style flexibility, PocketPal AI still has the larger model catalog.
Google AI Edge Gallery: Google's Free Gemma 4 App
Google AI Edge Gallery is Google's free app for running its Gemma 4 models on the phone, and it is available on the iPhone App Store. The listing shows AI Chat with a thinking mode, agent skills, image and audio input, Prompt Lab, and model benchmarking. Google says the app is in active development. Full review: Google AI Edge Gallery Review.
- What it is: Google's on-device AI showcase app, built on Google's LiteRT-LM runtime. Free; no in-app purchases are listed.
- Install: App Store → "Google AI Edge Gallery". Requires iOS 17.0 or later. Models download separately inside the app.
- Version and last update: v1.0.13, released 8 October 2026 (App Store version history). Gemma 4 support arrived in v1.0.2 on 3 April 2026 — no stale warning.
- Models: the Gemma 4 family (E2B and E4B are the phone-sized variants; Google lists the LiteRT-LM files at about 2.58 GB and 3.65 GB), EmbeddingGemma 2, and FunctionGemma 270M. Models in LiteRT-LM format can also be imported from Hugging Face.
- Features: Thinking Mode (Gemma 4 models only), Agent Skills (for example Wikipedia lookup and maps, plus community skills), Ask Image, Audio Scribe, Prompt Lab, and model benchmarking.
- Limits: only LiteRT-LM models, not GGUF or MLX; the listing does not mention Shortcuts or Siri; PromptQuorum has no speed figures for it.
- Best for: users who want Google's own Gemma 4 with a visible reasoning mode and agent tools, at no cost.
💡Tip: Google reports peak memory of about 607 MB on CPU and 1,450 MB on GPU for Gemma 4 E2B on an iPhone 17 Pro (vendor figures, per Google's LiteRT-LM Gemma 4 page; not re-measured by PromptQuorum). E2B is the practical starting point for most phones; E4B suits 8 GB+ iPhones.
MLC Chat: Apple Silicon Optimisation
Stale warning: the MLC Chat iPhone app has had no update since version 1.6 on 27 September 2024 (more than two years), so it may be unmaintained — check before relying on it. The MLC LLM engine repository itself is still active on GitHub (last activity 6 October 2026). MLC Chat is free, open-source, and runs models compiled by the MLC LLM toolchain rather than standard GGUF, and its catalog has not kept pace with PocketPal AI or Locally AI. Full review: MLC Chat Review.
- What it is: the iOS reference app from the MLC LLM project, demonstrating MLC LLM's Metal-accelerated inference on Apple Silicon.
- Install: App Store → "MLC Chat". Free. Requires iOS/iPadOS 17.0 or later.
- Version and last update: v1.6, released 27 September 2024 (App Store version history; added Llama 3.2 3B). Earlier: 1.5 added Gemma 2 2B (August 2024), 1.4 added Phi-3, Mistral v3 and Qwen2 (June 2024). Last iPhone update more than two years ago — stale warning on.
- Speed (project claim, not measured here): the MLC LLM project describes its Metal-accelerated engine as faster than llama.cpp-based apps on the same iPhone — but the app's own model list predates current picks like Phi-4 Mini and Gemma 3, so this advantage is untestable on 2026's recommended models without a self-compiled build.
- Model library: limited to models the MLC LLM project compiled before the app went dormant — Llama 3.2 3B, Gemma 2 2B, Phi-3, Mistral v3, Qwen2 and similar 2024-era releases. Not every Hugging Face GGUF works, and no current-generation small model ships in the App Store build.
- Best for: developers already invested in the MLC LLM toolchain who are willing to compile their own model, or who specifically need one of the pre-2025 models it already ships.
⚠️Warning: Because the App Store build of MLC Chat has had no update since 27 September 2024, do not expect Phi-4 Mini, Qwen3, or Gemma 3 to appear in its picker. If you want current-generation models with Metal-class GPU efficiency today, Locally AI (Apple MLX), PocketPal AI (llama.cpp with Metal Performance Shaders), or Google AI Edge Gallery (Gemma 4) are the actively maintained alternatives.
LLM Farm: Not on the App Store, Source Still Available
Stale warning: LLM Farm's last release (1.4.3) was on 27 January 2025 and its last code push was on 30 January 2026 — nearly nine months with no update — so it may be unmaintained. PromptQuorum's review of the project dates its removal from the App Store to August 2025; its own GitHub README describes the app as "temporarily unavailable" on both the App Store and TestFlight. The project (GitHub: guinmoon/LLMFarm) is still open-source and has seen GitHub issue activity into October 2026, but new users can no longer install it from the App Store; it must be built from source via Xcode (README: macOS 13+ and iOS 16+). Full review: LLM Farm Review.
- What it is: an iOS app from the project's maintainer that runs GGUF models with extensive configuration options, previously distributed on the App Store.
- Current availability: not installable from the App Store or TestFlight as of this refresh. Source code remains public and buildable.
- Building from source: clone the repository, open in Xcode, and build to a device with a free or paid Apple Developer account — the standard sideloading workflow for any open-source iOS app without an App Store listing.
- Configuration exposed (when built): temperature, top-p, top-k, mirostat sampling, repeat penalty, system prompt per model, chat template selection, context window length.
- Best for: developers comfortable building an app from source in Xcode who specifically want mirostat sampling controls. Everyone else should use PocketPal AI or Locally AI, both fully App-Store-installable and free.
⚠️Warning: Do not expect an App Store link for LLM Farm to work — treat any such link found elsewhere as outdated. Building from source requires Xcode and basic familiarity with iOS code signing. If that is more setup than you want, PocketPal AI covers the same "flexible free chat app" use case without leaving the App Store.
Apple Intelligence: System-Integrated On-Device AI
Apple Intelligence runs Apple's own ~3B foundation model on-device on iPhone 15 Pro and newer (the iPhone 15 Pro generation, reported at 8 GB RAM; Apple does not state RAM). It is not a chat app by default — it powers system features in Mail (Smart Reply), Messages (writing tools), Notes (summarisation), and Notification Summaries. Since iOS 26, Apple's Foundation Models framework lets third-party developers call that same on-device model directly, and apps like Locally AI now use it to expose a genuine chat interface on top of it. iOS 27 extends the framework: Apple's developer guide says apps can use any language model that conforms to its language-model protocol, including Apple's own models and cloud models.
- Where it lives: integrated into iOS 18+ (Foundation Models framework added in iOS 26). Enable in Settings → Apple Intelligence & Siri.
- Hardware requirement: iPhone 15 Pro / 15 Pro Max, iPhone 16 series, iPhone 16e, iPhone 17 series, and the 2026 iPhone 18 Pro, 18 Pro Max, and iPhone Duo (A20 Pro). Earlier iPhones (14 and below) do not support Apple Intelligence. Check Apple's current compatibility list for iOS 27 before buying.
- Siri AI (iOS 27): Apple describes Siri AI as available in beta in iOS 27, rolling out in English first, not initially available in the EU on iOS, and subject to daily usage limits for some server-side features. It is separate from the on-device chat apps in this guide.
- On-device features: Writing Tools (rewrite, summarise, proofread) inside any text field, Smart Reply in Mail and Messages, Notification Summaries, Genmoji generation.
- Foundation Models framework (iOS 26+): a native Swift API giving developers direct access to the same on-device model with a few lines of code — this is how Locally AI surfaces it as a chat option, and how the Shortcuts "Use Model" action can route a prompt to it, to Private Cloud Compute, or to ChatGPT.
- Private Cloud Compute: for tasks beyond the on-device model's capability, Apple Intelligence can send the request to Private Cloud Compute (PCC) — Apple-operated servers running larger models which Apple states do not retain user data. See Apple's Apple Intelligence & Privacy documentation for the controls available on your device.
- How it relates to chat apps: Apple Intelligence is a complement, not a replacement. Its system features handle text rewriting and summarisation inside iOS apps; PocketPal AI, Private LLM, Locally AI, and MLC Chat provide a dedicated chat interface for arbitrary questions — and Locally AI can now put Apple's own model behind that same kind of interface.
💡Tip: If Apple Intelligence is your only need (rewriting emails, summarising notifications), you do not need a separate chat app. If you want to ask the model questions like "explain quantum tunnelling in simple terms" or "draft a project plan for X", install a chat app — either a dedicated one like PocketPal AI, or Locally AI, which can front Apple's own on-device model with a chat window.
MemoryAI: Offline Chat with Editable Memory (Not Benchmarked)
MemoryAI is a free offline chat app for iPhone with document Q&A, an editable on-device memory and a PIN-locked Kid Mode. It was added after the main comparison, so it is not in the comparison table and was not part of our testing.
- Model library: Models download inside the app (0.3–4 GB each), and custom GGUF files can be imported from Hugging Face. It uses two inference engines: llama.cpp and Google's LiteRT.
- Features: Document Q&A, an Agent Memory you can open, edit and delete, hands-free voice (on-device speech input and read-aloud), image understanding with vision models, and a Kid Mode locked behind a parent PIN.
- Requirements and price: iOS 16.4+, 4 GB+ RAM. Free, with no ads, subscriptions or in-app purchases.
- Privacy and licence: Chat, documents and memories stay on the device after the model download. The app sends usage events and crash reports (no chat content, per the developer), and it is closed source.
- Install path: App Store. First released on 30 March 2026, so it is a young app. The App Store lists version 1.0.60.
💡Tip: The details above come from the store listings, the privacy policy and the developer, who reviewed our write-up; we have not timed MemoryAI on our test phones. Full notes: MemoryAI review.
fullmoon: Free Open-Source Chat on Apple MLX (No Release Since Early 2025)
fullmoon is a free, MIT-licensed offline chat app for iPhone, iPad and Mac that runs small models through Apple's MLX, but it has had no release since 28 January 2025 and no code change since 5 May 2025, so it may be unmaintained. It was added after the main comparison, so it was not part of our testing.
- Models: Llama 3.2 1B and 3B and two DeepSeek-R1-Distill-Qwen-1.5B builds (0.7–1.9 GB each), per the website.
- Features: An editable system prompt, theme and fonts, and a Shortcuts action; chat history is stored locally (README).
- Requirements and price: iOS 17.6+, 31.6 MB before models are downloaded. Free to download.
- Privacy and licence: The policy says nothing is collected but also lists purposes for collected data, so it contradicts itself. The source code is public under the MIT licence.
- Install path: App Store (version 1.2.0); the website also offers a TestFlight link for newer builds.
💡Tip: The details above come from the maker's website, the App Store listing, the privacy policy and the public repository; we have not run fullmoon on our test phones. Full notes: fullmoon review.
Cloak: Local AI: Paid-Once Offline Chat (Not Benchmarked)
Cloak: Local AI is a paid-once offline chat app for iPhone and iPad with Voice Memos and a weekly Recap. It was added after the main comparison, so it is not in the speed figures and was not part of our testing.
- Model library: The app downloads one recommended model from Hugging Face during setup; there is no model picker, and the model name and size are not published.
- Features: Multiple local chat threads, a customizable system prompt, Voice Memos turned into notes or chats, and Recap, a summary of the last seven days built from Photos, Location and Voice Memos. A separate Mac version, Cloak: Local AI Chat, is chat only.
- Requirements and price: iOS or iPadOS 18.0+. A one-time purchase of $1.99 in the US store, with no subscription and no in-app purchases listed.
- Privacy and licence: The policy says conversations stay on the device after the model download. The iPhone version sends an optional setup survey answer and, only if switched on, daily feature-use counts; it is closed source.
- Install path: App Store. First released on 25 June 2026; the App Store lists version 1.2.9, released on 27 October 2026.
💡Tip: The details above come from the App Store listings, Polyaura's website and its privacy policy; the developer was not contacted, and we have not timed Cloak on our test phones. Full notes: Cloak: Local AI review.
Reins: Chat Client for Your Ollama or LM Studio Server (Not Benchmarked)
Reins is a free chat app for iPhone and iPad that connects to your own Ollama, LM Studio or OpenAI-compatible server and, since version 2, can also run MLX models on the device. It was added after the main comparison, so it is not in the comparison table and was not part of our testing.
- Model library: On-device MLX models download inside the app from Hugging Face, and any MLX model can be searched and downloaded; Apple Foundation Models are supported too.
- Features: Tool calls, built-in web search without an API key, image, PDF, CSV and text attachments, chat branching, thinking mode and temporary chats.
- Requirements and price: iOS 17.0+ (background generation needs iOS 26). Free to download, with Reins Pro as a monthly, yearly or lifetime in-app purchase; the sources do not list what Pro unlocks.
- Privacy and licence: The developer says there is no login, no analytics and no data collection. Version 2 is closed source; the public repository holds the code only up to version 1.3.4.
- Install path: App Store. The App Store lists version 2.3.3, released on 21 October 2026.
💡Tip: The details above come from the App Store listing, the developer's website and policies; we have not timed Reins on our test phones. Full notes: Reins review.
Models by iPhone Tier
iPhone RAM determines the model size ceiling — not the chip generation. A 6 GB iPhone (14 Pro, 15) can comfortably run 1.7B models; an 8 GB+ iPhone (15 Pro, 16 series, 16e, 17 series) runs 3B–4B models comfortably and 7B models slowly. Newer small models worth trying alongside Phi-4 Mini: Gemma 4 E2B and E4B (Google, April 2026; about 2.58 GB and 3.65 GB as LiteRT-LM files) and the Qwen3.5 small series (Alibaba, March 2026; 0.8B, 2B, 4B, and 9B, Apache 2.0 per the Artificial Analysis summary). PromptQuorum has no speed figures for them; Locally AI and Google AI Edge Gallery ship them directly. For the broader model landscape across all hardware (not just mobile), see Best Local LLMs in 2026.
iPhone Tier (Year, RAM) | Recommended Model | Download Size | Rough speed (indicative) |
|---|---|---|---|
| iPhone 18 Pro / Pro Max (2026, ~12 GB reported) | Start as 17 Pro; RAM not published by Apple | ~2.5–3.7 GB | No figure |
| iPhone 17 Pro / 17 Pro Max (2025, 12 GB) | Phi-4 Mini / Llama 3.2 3B / Gemma 4 E4B | ~2.5–3.7 GB | ~13–20 tok/sec (Phi-4 Mini) |
| iPhone 16 series / 16e / 17 (2024–2025, 8 GB) | Phi-4 Mini (3.8B Q4_K_M) or Gemma 4 E2B | ~2.6–2.7 GB | ~10–15 tok/sec (Phi-4 Mini) |
| iPhone 15 Pro / Pro Max (2023, 8 GB) | Phi-4 Mini (3.8B Q4_K_M) | ~2.7 GB | ~8–12 tok/sec |
| iPhone 13 Pro / 14 Pro / Pro Max (6 GB) | Qwen3 1.7B or SmolLM 2 1.7B (Q4_K_M) | ~1.1 GB | ~15–20 tok/sec |
| iPhone 14 / 14 Plus / 15 / 15 Plus (6 GB) | Qwen3 1.7B or SmolLM 2 1.7B (Q4_K_M) | ~1.1 GB | ~12–18 tok/sec |
| iPhone 13 / 12 / SE / older (4 GB) | Not recommended for on-device LLM | — | — |
💡Tip: For older 6 GB iPhones, Qwen3 1.7B or SmolLM 2 1.7B (HuggingFace) are sensible choices in 2026; Qwen3.5 2B and Gemma 4 E2B are newer options in the same size class. Models this small produce coherent short-form responses (1–3 paragraphs) but struggle with multi-step reasoning. Do not install Phi-4 Mini on a 6 GB iPhone — it nominally fits but iOS will kill the app under any other memory pressure. Speed figures in the table are rounded indicative ranges for Phi-4 Mini, Llama 3.2 3B, and Qwen3 1.7B, not tied to an app version or test date; Gemma 4 and Qwen3.5 have no figures here. Apple's product pages do not state iPhone RAM: the RAM figures here are widely reported specs, and the ~12 GB for the 18 Pro comes from press reports.
Battery Drain and Thermal Throttling
On-device LLM inference on iPhone is CPU/GPU-intensive and produces heat. As a rough guide (rough orientation, not a measurement by PromptQuorum), active inference draws ~3–5 W; sustained generation will throttle the chip and can drain the battery roughly 20–30% per hour on iPhone 16 Pro.
- Battery drain (active chat, indicative): ~20–30% per hour on iPhone 16 Pro running Phi-4 Mini.
- Sustained generation can trigger thermal throttling. When the chip gets hot, iOS reduces clock speeds and tokens/sec drop. When it starts and how much speed you lose depends on the phone, case, charging, and room temperature; PromptQuorum has not measured this. Letting the phone cool restores speed.
- Mitigation: keep the iPhone face-up on a hard surface (not in your hand or pocket) during long inference sessions to allow heat dissipation. A passive heat sink case helps but is rarely necessary for short interactions.
- Phantom drain: if you leave a chat app open in the background after generation, RAM stays allocated but no inference runs — battery impact is minimal. Closing the app fully reclaims the ~3 GB of RAM.
- MagSafe charging during inference: charging while generating can raise heat; if the phone gets warm, pause charging or generation, or charge afterwards.
⚠️Warning: Do not run on-device LLM inference on an iPhone in direct sunlight or in a hot car. In direct sunlight or a hot car the chip can reach its thermal limits sooner, which can trigger throttling and Apple's "iPhone needs to cool down" warning. Generation usually slows rather than the app closing.
iOS Shortcuts, Siri, and Sideloading
iOS integration varies significantly by app. Private LLM and Locally AI both support Shortcuts; PocketPal AI, Google AI Edge Gallery, and MLC Chat are standalone chat apps with no Shortcuts or Siri integration mentioned in their README or App Store listing.
Private LLM Shortcut: summarise selected text
“1. Action: "Get Selected Text" (iOS Share Sheet input). 2. Action: Private LLM text generation (from the Private LLM app's Shortcuts actions) → Prompt: "Summarise the following text in three bullet points: [Selected Text]" → pick a model installed in the app. 3. Action: "Show Result" or "Copy to Clipboard". Add to Share Sheet so you can run it on any selected text in any app, without WiFi once the model is downloaded. Action names follow the developer's description and may differ by app version.”
Apple Intelligence Shortcut: rewrite for tone
“1. Action: "Get Clipboard". 2. Action: "Use Model" → Model: On-device → Prompt: "Rewrite this in a professional, concise tone: [Clipboard]". 3. Action: "Copy to Clipboard". Assign to a Lock Screen widget for one-tap rewriting of anything you copy.”
- Private LLM plugs into Shortcuts and Siri, per the developer, and can hand responses to apps that support x-callback-url. The most iOS-native of the chat apps here.
- Locally AI exposes a Shortcuts action for free, and can route a Shortcut to Apple's own on-device foundation model via its Foundation Models framework integration — no Siri voice trigger, but Shortcuts automation without paying.
- PocketPal AI is a standalone chat app — its README mentions no Shortcuts action and no Siri integration. You open the app and chat.
- Google AI Edge Gallery is a standalone chat and demo app — its App Store listing mentions no Shortcuts action or Siri integration.
- MLC Chat is a reference app for the MLC LLM project — minimal iOS integration and no iPhone update since September 2024. No Shortcuts action.
- LLM Farm is currently not on the App Store and, when built from source, has no Shortcuts action.
- Apple Intelligence integrates with iOS Shortcuts via the "Use Model" action (iOS 18.4+, expanded in iOS 26 alongside the Foundation Models framework). This routes a prompt to the on-device model, Private Cloud Compute, or ChatGPT (configurable). The on-device output can be chained into other Shortcuts actions.
- Sideloading: PocketPal AI, Private LLM, Locally AI, Google AI Edge Gallery, and MLC Chat are on the App Store and require no sideloading or jailbreak; Apple Intelligence is built into iOS. LLM Farm is the exception — installing it now requires building from source in Xcode. In the EU, alternative app marketplaces exist under the DMA, but which apps they offer is up to each developer, so install from the developer's official channel; none of these apps needs one.
💡Tip: For hands-free use while driving or cooking, Private LLM's Siri integration is the only chat app in this guide whose developer lists Siri integration, which lets you use it without touching the phone. Apple Intelligence supports voice via Siri but only for system tasks (writing, summarising, app actions) — it does not expose general Q&A like the chat apps do.
Common Mistakes
- Installing a model larger than your iPhone's RAM allows. A 7B model on an 8 GB iPhone is often too slow for chat and can be closed by iOS when it reclaims memory for another app; measure it with PocketPal AI's built-in benchmark first. Stick to the recommended model for your iPhone tier (3B–4B for 8 GB devices, 1.7B for 6 GB devices).
- Expecting cloud-AI quality from on-device models. Phi-4 Mini (3.8B) is impressive for its size but not GPT-5.5. Use it for chat, summarisation, drafting, and quick Q&A — not multi-step reasoning, complex code generation, or nuanced creative writing.
- Running inference in direct sunlight or a hot car. Thermal throttling can kick in quickly, generation slows, and you may see the "iPhone needs to cool down" warning. Run inference at room temperature.
- Leaving 3+ chat apps installed with 3 GB models in each. You will burn ~10 GB of storage on duplicate models. Pick one app and one model; uninstall the others until you are sure you need them.
- Assuming Apple Intelligence has no chat interface at all. That was true through iOS 18, but since iOS 26 apps like Locally AI can front Apple's on-device model with an actual chat window via the Foundation Models framework. The system features (Writing Tools, Smart Reply) still are not a chat UI on their own.
- Looking for LLM Farm on the App Store. It is currently not on the App Store (PromptQuorum's LLM Farm review dates the removal to August 2025). Searching the App Store for it wastes time; use PocketPal AI for the same free/flexible use case, or build LLM Farm from source in Xcode if you specifically need its exposed mirostat controls.
- Assuming "free" means no purchases. PocketPal AI is free for local chat but sells optional premium content (Pals and "PocketPal Plus"), and Locally AI and Google AI Edge Gallery list no in-app purchases. Check the listing's in-app purchase section before you install (PocketPal AI's US App Store listing showed an 18+ age rating on 2026-10-23).
- Judging a model from a speed table alone. The tokens/sec ranges in this guide are indicative, and newer models such as Gemma 4 and Qwen3.5 are not in them. Run the same prompt on two models in PocketPal AI's benchmark before you delete the one you already downloaded.
Sources
- PocketPal AI — github.com/a-ghorbani/pocketpal-ai (README, MIT license, release v1.20.1 of 19 Oct 2026) and the App Store listing (version history, in-app purchases, age rating, iOS 15.1+). Read 2026-10-23.
- Private LLM — App Store listing (v1.9.16, $4.99, iOS 17.0+, Family Sharing) and privatellm.app (Siri and Shortcuts, model list, one purchase for iPhone, iPad, and Mac). Read 2026-10-23.
- Locally AI — App Store listing (v1.66.0, free, Element Labs Inc, iOS 18.1+; LM Studio team) and lmstudio.ai/locally (LM Link). Read 2026-10-23.
- Google AI Edge Gallery — App Store listing (v1.0.13, free, iOS 17.0+, Gemma 4, Agent Skills, Thinking Mode), github.com/google-ai-edge/gallery and Google's LiteRT-LM Gemma 4 page (E2B/E4B file sizes and memory, vendor figures). Read 2026-10-23.
- MLC Chat / MLC LLM project — App Store listing (v1.6, 27 Sep 2024, version history), llm.mlc.ai and github.com/mlc-ai/mlc-llm (engine repository). Read 2026-10-23.
- LLM Farm — github.com/guinmoon/LLMFarm (README: "temporarily unavailable" in TestFlight and App Store; release 1.4.3 of 27 Jan 2025; last push 30 Jan 2026) and the PromptQuorum LLM Farm review for the August 2025 delisting. Read 2026-10-23.
- Apple Intelligence and on-device foundation models — Apple Machine Learning Research, the What's new in iOS 27 developer guide (Foundation Models framework), and the iPhone 18 Pro page (A20 Pro, Dual 16-core Neural Engine, Siri AI beta, no RAM figure) and iPhone Duo page. Read 2026-10-23.
- Qwen3.5 small series (0.8B–9B, March 2026) — Artificial Analysis summary.
Frequently Asked Questions
Can iPhone really run a 7B model?
Technically yes on iPhone 15 Pro and newer (8 GB RAM), but not at usable speed. A 7B Q4 model on iPhone 16 Pro is often too slow for comfortable chat (indicative). iOS also tends to kill the app when other apps need memory. Use 3B–4B models (Phi-4 Mini, Llama 3.2 3B, Gemma 3 4B) for everyday on-device chat. For 7B+ quality, remote-connect to a home Mac or PC running Ollama.
Does local AI drain my iPhone battery?
Yes — actively running inference typically draws ~3–5 W and, as a rough indicative guide, can drain the battery roughly 20–30% per hour on iPhone 16 Pro. For occasional chat (a few prompts), the impact is small. For sustained use (long conversation, multiple summarisation tasks), expect the battery to drain quickly; charging while generating can add heat, so pause generation if the phone gets warm. The model itself sitting in RAM with no inference running has negligible battery impact.
Will my iPhone get hot using local AI?
Yes, it can, especially during long continuous generation. iOS throttles clock speeds when the chip gets hot, which lowers tokens/sec; how soon and how much varies with the phone, case, charging, and room temperature, and PromptQuorum has not measured it. To minimise this: keep the iPhone face-up on a hard surface (not in your hand) during long sessions, and avoid direct sunlight. Short interactions usually cause little noticeable heat.
Can I use Siri with a local model?
Yes, with Private LLM ($4.99 one-time). The developer states it plugs directly into Siri and Shortcuts, so a prompt can be routed to the on-device model without an internet connection. PocketPal AI, Locally AI, Google AI Edge Gallery, and MLC Chat do not mention Siri voice integration, though Locally AI does support Shortcuts automation. Apple Intelligence integrates with Siri but only for system tasks (writing, summarising, app actions), not general Q&A; Siri AI is a separate iOS 27 beta feature, English first and not initially in the EU.
Do these apps work on iPhone SE or older iPhones?
Limited. iPhone SE and iPhone 13 / 12 (4 GB RAM) are below the practical threshold for on-device LLMs in 2026. iPhone 14 / 15 (non-Pro, 6 GB RAM) can run 1.7B models (Qwen3 1.7B, SmolLM 2 1.7B) but not 3B+. iPhone 13 Pro and 14 Pro (6 GB RAM) also stay at 1.7B; the iPhone 15 Pro (8 GB RAM) can run 3B models like Phi-4 Mini at roughly 8–12 tokens/sec (indicative). For older iPhones, the better path is remote-connecting to a home Mac or PC.
Can I sync chat history between iPhone and Mac?
Not in PocketPal AI, Private LLM, Google AI Edge Gallery, or MLC Chat — chat history is stored locally per device, with no iCloud sync. Private LLM and Locally AI are both universal purchases/downloads that run natively on Mac too, but history still does not sync between the iPhone and Mac copies. For cross-device chat history, the practical approach is to run Open WebUI on a home Mac and access it from both iPhone and Mac browsers — Open WebUI stores chat history server-side.
Are these apps available outside the App Store?
PocketPal AI is open-source and can be built from source via Xcode, but the App Store build is the standard distribution. LLM Farm is the exception that now requires building from source — it is currently not on the App Store or TestFlight (the README says "temporarily unavailable"). Private LLM, Locally AI, Google AI Edge Gallery, and MLC Chat are App Store only. In the EU, alternative app marketplaces exist under the DMA, but which apps they offer is up to each developer, so install from the developer's official channel; none of these apps needs one.
Do any of them require jailbreak?
No. PocketPal AI, Private LLM, Locally AI, Google AI Edge Gallery, MLC Chat, and Apple Intelligence all run on stock iOS. Building LLM Farm from source in Xcode also does not require jailbreak — it uses standard developer sideloading, not a jailbreak exploit. Jailbreaking is not required and not recommended for any of these apps.
Can I use local AI in iOS Shortcuts?
Yes, via Private LLM (its Shortcuts actions), Locally AI (a free Shortcuts action), or Apple Intelligence (the "Use Model" action, iOS 18.4+, expanded in iOS 26). PocketPal AI, Google AI Edge Gallery, and MLC Chat do not mention Shortcuts actions. Among the apps compared here, Locally AI is the only free option that lists both Shortcuts support and access to Apple's own on-device model.
How does local AI compare to the ChatGPT app on iPhone?
On-device models (Phi-4 Mini, Llama 3.2 3B) remain meaningfully behind frontier cloud models for complex reasoning, broad world knowledge, and multimodal tasks, but are faster on simple queries (no network round-trip) and, per each app's documentation, keep chat on the device. The honest tradeoff: local AI for routine and private tasks; ChatGPT or another cloud app for occasional hard questions. Many users in 2026 keep both and choose per query.
Is LLM Farm still safe to use if it is no longer on the App Store?
The source code is public on GitHub and still receives issue activity into 2026, but the last release (1.4.3) was on 27 January 2025 and the last code push on 30 January 2026, so it may be unmaintained. Built from source and sideloaded, it is also no longer reviewed by Apple's App Store process. If you do not specifically need its mirostat sampling controls, PocketPal AI covers the same free/flexible/open-source use case while staying on the App Store with Apple's standard app review.
Which of these iPhone apps are free?
PocketPal AI is free (MIT license) with optional in-app purchases for premium Pals and "PocketPal Plus". Locally AI and Google AI Edge Gallery are free with no in-app purchases listed. MLC Chat is free. LLM Farm is free but must be built from source. Private LLM is $4.99 one-time with no in-app purchases listed. Apple Intelligence is built into iOS. Prices were read from the US App Store on 2026-10-23 and can differ by country.
Can the iPhone Duo or iPhone 18 Pro run a local LLM offline?
In principle yes: both use Apple's A20 Pro chip, and the apps in this guide run on any iPhone that meets their iOS minimum. What decides the model size is RAM, and Apple's iPhone 18 Pro and iPhone Duo pages do not state a RAM figure; press reports put the iPhone 18 Pro at 12 GB, which is unconfirmed. PromptQuorum has no measured speeds for either phone. Install PocketPal AI and run its built-in benchmark to see what your phone actually does.
Do these apps work on iPad?
Yes for most. PocketPal AI (iPadOS 15.1+, full iPad support per its README), Private LLM (iPadOS 17.0+), and Locally AI (iPadOS 18.1+) list iPad support. The Google AI Edge Gallery listing requires iOS 17.0 or later; check its page for your iPad. As on iPhone, RAM decides which models fit. For iPad-specific advice see How to Run AI on a Tablet.
Why did LLM Farm disappear from the App Store?
The project README says: "The app is temporarily unavailable in TestFlight and Appstore." PromptQuorum could not find a public reason for the removal, so any explanation you read elsewhere is unconfirmed (see the LLM Farm review). The source code is still public, but the last release was 1.4.3 on 27 January 2025. For a free, App Store-installable alternative, use PocketPal AI or Locally AI.
Is there an Ollama app for iPhone?
Ollama's download page lists macOS, Windows, and Linux, and no iPhone app. To use Ollama or LM Studio models from an iPhone, run the server on a Mac or PC and connect from a client app; Locally AI offers LM Link, which LM Studio describes as an encrypted connection to LM Studio on a Mac. For models that run entirely on the phone, use the apps in this guide.
Should I download Gemma 4 or Phi-4 Mini?
PromptQuorum has no head-to-head benchmark of the two on iPhone. Facts that help: Gemma 4 E2B and E4B are Google's phone-sized models (about 2.58 GB and 3.65 GB as LiteRT-LM files in Google AI Edge Gallery; Locally AI also lists Gemma 4), while Phi-4 Mini (3.8B Q4_K_M, ~2.7 GB) is the model the speed figures in this guide use. On an 8 GB iPhone try both against your own prompts and compare tokens/sec in PocketPal AI's benchmark; on a 6 GB iPhone stay with 1.7B-class models.