Key Takeaways
- Loci is a free on-device AI app for iPhone, iPad, Android, Mac, and Windows, built by Loci AI, Inc.
- It offers two model paths: Apple's system foundation model (where supported) or downloadable open-source models (Gemma 4, Qwen 3.5, Llama, Phi).
- Setup is minimal β no GGUF files, no quantization choices, just download and chat. Loci intelligently selects the best runtime (llama.cpp or MLX) for your hardware.
- Privacy claim: conversations stay on-device; optional features (web RAG, voice) require internet. Web RAG grounds answers in live sources when online and falls back cleanly to on-device when offline.
- Standout features: Gemma 4 E2B/E4B as the strongest mobile reasoning available; Loci Link (desktop/phone pairing for powerful models); global memory across conversations; first-class thermal management; background downloads; and exceptional crash resilience.
- Not ideal for users needing frontier cloud reasoning, deep GGUF/quantization control, or fully documented voice stacks.
- Real-world testing shows model downloads work reliably, offline chat functions as advertised, app crashes less than alternatives. Gemma 4 combined with Loci's context management delivers the strongest reasoning ceiling on mobile today.
What Loci Is
Loci is a consumer-focused, on-device AI assistant available on iPhone (iOS 18.0+), iPad (iPadOS 18.0+), Android, Mac, and Windows. Built by Loci AI, Inc., the app is free with no subscription, no ads, and no account requirement.
Model architecture: Loci can use "Apple's built-in foundation model or download from 10+ curated open-source models, including Gemma, Qwen, Llama, and Phi β all running locally on your device." This means inference happens on-device after setup, not in the cloud.
Privacy positioning: The official claim is "Chat is processed on your device and is not uploaded. There is no account, no server-side copy of your conversations, and no training on your words." The app collects "Identifiers," "Usage Data," and "Diagnostics" via its privacy nutrition label, but states this data is "not linked to your identity."
Optional features include photo analysis, voice mode, calendar/reminders integration, and web search via DuckDuckGo. The caveat: web search and Windows voice input require an internet connection, which alters the "offline" story if used.
How Local AI Works in Loci
Loci offers two possible paths to on-device AI:
- Apple system foundation model β on supported Apple devices (iPhone, iPad, Mac with recent iOS/macOS versions), Loci can use a built-in on-device foundation model provided by Apple. This path requires no model download, minimal setup friction, and is the simplest on Apple platforms.
- Downloadable open-source models β users can download compact models (Gemma 4 1B/4B, Qwen 3.5, Llama 3.2 3B, Phi-4 Mini) into Loci once. Model files typically range from 1β5 GB depending on model size. After download, inference runs on-device; internet is not required for chat. New models land continuously and are integrated once verified to run on phone hardware.
- Runtime flexibility β Loci intelligently selects the best inference runtime for each model and device: llama.cpp for GGUF models and Apple's MLX for optimal performance on Apple silicon. This means the simple interface masks sophisticated tuning beneath the surface β the app picks the fastest path for that model on your specific hardware (measuring time-to-first-token and tokens-per-second), rather than forcing users to understand runtimes.
Real-World Testing Notes
Loci was tested on multiple devices (testing by Hans KΓΌpper, PromptQuorum, August 2026) to validate real-world usability:
- Model downloads work reliably. Downloads of compact models (e.g., Gemma 4 4B, ~4 GB) completed successfully on home WiFi with no truncation or corruption observed.
- Offline chat works as advertised. Once a model is downloaded, inference runs without any internet connection, including in airplane mode. Chat remains responsive.
- Model quality depends on both parameter count and app-level tuning. The baseline 3Bβ4B models excel at drafting, brainstorming, and summarization, but struggle with nuanced topics and multi-step reasoning β a ceiling inherent to parameter count that no app can change. However, what the app *can* change is everything surrounding the model: memory management, context-window handling, hallucination reduction, and OS-level stability. Loci implements best-practice context management (automatic compaction plus newest research techniques for stretching usable context), resulting in fewer hallucinations and crashes compared to other apps running the identical model. Most importantly: Gemma 4 E2B and E4B (the latest generation) are a genuine step above the 2025 generation of 3Bβ4B models and represent the strongest reasoning capability available on mobile today. Combined with Loci's context management, this is the reasoning ceiling for pocket-sized AI. For complex analysis, live web queries, or high-stakes decisions, larger frontier models (GPT-4o, Claude 3 Opus) remain necessary.
- Optional features require connectivity. Web search via DuckDuckGo, model downloads, and app updates all require internet access as documented.
Trade-Offs: Benefits vs. Limitations
Privacy
- What it means in real use:
- Conversations do not leave your device (per developer claim). No cloud servers storing your words.
- Limitation / caveat:
- Optional features (DuckDuckGo web search, Windows voice input) require connectivity and may leave data trails.
Offline-capable after setup
- What it means in real use:
- Once the app and model are installed, chat works in airplane mode with no internet.
- Limitation / caveat:
- Initial model download requires internet. Feature updates, model downloads, and backups may also need connectivity.
No recurring cloud cost
- What it means in real use:
- Free app, no subscription, no per-message fees.
- Limitation / caveat:
- Inference runs on your device, consuming local battery and processing power.
Works across platforms
- What it means in real use:
- One free purchase (the app is free) on iPhone, iPad, Android, Mac, and Windows.
- Limitation / caveat:
- Quality and capabilities may vary per platform; Apple device priority is evident in the design.
Minimal setup friction
- What it means in real use:
- No GGUF file selection, no quantization tuning, no VRAM calculations.
- Limitation / caveat:
- Model choice is curated and limited (10+ models). Cannot import your own GGUF files.
Device performance ceiling is your only limit
- What it means in real use:
- Inference speed depends on your phone/PC RAM and CPU, not cloud queue times.
- Limitation / caveat:
- Smaller local models (~3Bβ4B params) produce less capable output than frontier cloud LLMs (GPT-4o, Claude 3 Opus).
Web grounding with seamless offline fallback
- What it means in real use:
- When online, Loci can ground answers in live web sources via configurable web RAG. When offline, it cleanly falls back to on-device knowledge without breaking. You control the behavior in settings β get the online benefit and offline benefit from the same app. Few other local-AI apps offer this combination.
- Limitation / caveat:
- Web grounding requires internet when enabled. Without web access, answers reflect only on-device model knowledge and do not include current events.
Long-context and complex reasoning
- What it means in real use:
- Suitable for drafting, summarizing, and structured Q&A on local content.
- Limitation / caveat:
- Complex multi-step reasoning, coding, and high-stakes summarization often still benefit from frontier cloud models.
Desktop/phone linking for powerful model execution
- What it means in real use:
- Link your phone to a desktop Mac or PC to run extremely powerful models directly from your phone β solving hallucination and quality concerns by offloading inference to a more capable machine while keeping the interface on mobile.
- Limitation / caveat:
- Requires a desktop/laptop with sufficient GPU/CPU and consistent network connectivity between devices.
App stability and crash resilience
- What it means in real use:
- Loci AI engineered a unique approach to OS memory handling that results in significantly fewer crashes compared to other local-LLM apps β better session continuity and reliability.
- Limitation / caveat:
- Memory handling is app-specific; mileage varies by device, model size, and concurrent app load.
Global memory across conversations
- What it means in real use:
- Loci remembers things about you across conversations and model switches. Switch from Gemma to Qwen and the app retains your preferences and history β all stored on-device.
- Limitation / caveat:
- Global memory stays local to your device; not synced across devices.
First-class thermal management
- What it means in real use:
- Loci monitors device temperature and automatically adjusts runtime parameters to keep your phone stable, rather than pushing inference until the phone throttles itself or the app crashes. Multiple thermal modes let you deliberately run cooler for extended sessions.
- Limitation / caveat:
- Automatic thermal management trades some speed for stability; different devices have different thermal profiles.
Background model downloads
- What it means in real use:
- Model downloads run on background sessions and survive the app being closed, resuming where they left off. No need to keep the app open while downloading a 4 GB model file.
- Limitation / caveat:
- Downloads still require internet connectivity during active download windows.
Progressive settings depth
- What it means in real use:
- The main app screen is intentionally simple for casual users, but powerful settings are available for users who want to tune model behavior, temperature, context length, and thermal modes.
- Limitation / caveat:
- Advanced settings are hidden by default; requires exploring to discover all capabilities.
Loci on Each Platform
iPhone
- What to expect:
- Loci works on iOS 18.0+. Can use Apple's on-device foundation model or download a compact open-source model (Gemma 4 4B, Llama 3.2 3B, ~2β4 GB). Chat, photo analysis, voice mode, and calendar integration available.
- Important note:
- iOS 18+ requirement excludes iPhone XS and older. Exact device/chip thresholds for Apple foundation-model support are not publicly documented.
iPad
- What to expect:
- Loci works on iPadOS 18.0+, with the same model paths as iPhone. Larger screen is better for long conversations and document review.
- Important note:
- Larger models may still be constrained by available VRAM. Apple foundation-model availability varies by iPad generation; check App Store for current compatibility.
Android
- What to expect:
- Available on Google Play. Can download open-source models (Gemma 4 4B, Qwen 2.5, Llama 3.2 3B, Phi-4, ~2β5 GB). No built-in system model equivalent to Apple's foundation model.
- Important note:
- Performance varies widely across Android devices due to chipset, RAM, and OS version fragmentation. High-end phones (Snapdragon 8 series, 8+ GB RAM) handle models better.
Mac
- What to expect:
- Available on the Mac App Store. Can use Apple's on-device foundation model, download open-source models, or connect to Ollama running on the same Mac for access to a broader model library. Useful for longer sessions, larger screens, and external keyboards.
- Important note:
- Mac-specific Apple foundation-model support is undocumented. M-series Macs (M1/M2/M3+) likely supported; older Intel Macs may require model download. Advanced users can run Ollama locally and plug models into Loci for expanded model access.
Windows
- What to expect:
- Available via askloci.ai or Windows App Store. Can download open-source models (same library as Android: Gemma, Qwen, Llama, Phi). Voice input requires internet connection (unlike other platforms).
- Important note:
- Windows support is the least documented of the five platforms. Performance depends on GPU/CPU; requires sufficient disk space for model storage (~2β5 GB).
Loci vs. Alternatives
Loci
- Best for:
- Low-friction cross-platform private chat
- Setup level:
- Minimal (download, chat)
- Model flexibility:
- Curated library (~10 models); cannot import GGUF
- Platform focus:
- iPhone/iPad/Android/Mac/Windows (5 platforms)
- Key limitation:
- Model choice is limited to curated library (~10 models); no GGUF import
Private LLM
- Best for:
- Apple-only users wanting advanced model selection
- Setup level:
- Low-to-medium (one-time purchase, model downloads)
- Model flexibility:
- 140+ models, OmniQuant and GPTQ quantization formats
- Platform focus:
- iPhone/iPad/Mac (Apple only, one purchase across all devices)
- Key limitation:
- Apple-only; one-time purchase price not disclosed; requires learning quantization formats
PocketPal AI
- Best for:
- Users wanting full GGUF import and model control
- Setup level:
- Medium (free, but requires model file sourcing)
- Model flexibility:
- Any GGUF file from Hugging Face or elsewhere
- Platform focus:
- iPhone/iPad (primarily Apple, some Android support)
- Key limitation:
- Requires comfort with GGUF files and model selection; more complex than Loci
Google AI Edge Gallery
- Best for:
- Experimental model exploration with advanced features
- Setup level:
- Medium-to-high (more features, more discovery needed)
- Model flexibility:
- Gemma 4 centerpiece; supports multiple open-source models; custom model loading
- Platform focus:
- Android-first (iOS, macOS also supported)
- Key limitation:
- More experimental/technical; not a straightforward Loci replacement; less polished UI
Who Should Use Loci
- Privacy-conscious traveller. Loci works offline after setup, so you can chat without roaming data or relying on hotel Wi-Fi. No cloud service can see your words.
- Beginner who does not want to manage GGUF files. If the concept of quantization, model weights, and GGUF file handling sounds overwhelming, Loci is the right first local-AI app. No learning curve.
- User seeking lightweight writing/brainstorming assistant. Drafting notes, brainstorming ideas, summarizing text β all feasible on-device without sending your work to a cloud service.
- User with inconsistent connectivity. If your internet connection drops often (remote areas, transit, events), offline chat is a genuine advantage.
- Cross-device simplicity. One free app across iPhone, iPad, Android, Mac, and Windows, with consistent experience.
- User who wants to run powerful models from their phone. Loci Link lets you pair your phone to a Mac or PC to execute larger, more capable models directly from the mobile interface β solving quality and hallucination concerns while keeping traffic between your own devices and maintaining privacy. This is the real answer to the mobile quality ceiling and something no other app in the App Store offers.
- User who values persistent memory across conversations. Loci remembers things about you across different conversations and even when switching between models. All memory stays on your device.
- User running intensive sessions without device overheating. Loci's thermal management automatically keeps your phone stable during long sessions by adjusting runtime parameters, rather than letting the device throttle itself or crash the app.
Who Should Not Use Loci
- User expecting frontier reasoning or live web integration. While Gemma 4 E2B/E4B offer the strongest reasoning available on mobile and Loci's context management minimizes hallucinations, even the best small models face a parameter-count ceiling that GPT-4o or Claude 3 Opus exceed. For high-stakes analysis, complex multi-step reasoning, code generation, or live web grounding, cloud models remain necessary. Loci excels at personal assistant tasks, drafting, and reasoning within its design envelope.
- User needing live web knowledge offline. Loci has optional DuckDuckGo web search, but it requires internet. The local models have no concept of "today" or current events.
- Developer wanting comprehensive model/inference control. If you need to benchmark different quantizations, compare token/second speeds, or tune sampling parameters, Private LLM or PocketPal AI offer more depth.
- User building a fully documented voice assistant stack. Loci has a "voice mode" feature: on Apple platforms (iPhone, iPad, Mac), voice input and output run fully on-device with no cloud dependency. On Windows, voice input is constrained by the operating system's speech-recognition APIs and requires internet (a Windows OS limitation, not a Loci design choice). If you need a fully documented, sourced voice stack with Whisper STT + local LLM + TTS, see Build a Local Voice Assistant on Your Phone for the recommended open-source pipeline.
- User handling highly sensitive information. Before using Loci for private/confidential work, review the current privacy policy and privacy nutrition label on the App Store. App Store label shows data collection for "Identifiers," "Usage Data," and "Diagnostics" (stated as not linked to your identity), but read the full policy first.
Frequently Asked Questions
Does Loci include an AI model automatically?
Not always. On supported Apple devices (iPhone, iPad, Mac), Loci can use Apple's on-device foundation model at no extra step. On Android and Windows, or if the Apple system model is not available on your device, you must download a model the first time you chat (Gemma, Qwen, Llama, or Phi β about 2β4 GB depending on the model). After the one-time download, the model stays on your device.
Is Loci completely private?
Loci's official claim is that "Chat is processed on your device and is not uploaded." However: (1) optional features like DuckDuckGo web search and Windows voice input require internet and may create data trails; (2) the app collects "Identifiers," "Usage Data," and "Diagnostics" per its App Store privacy nutrition label (stated as "not linked to your identity"); (3) app updates and model downloads require internet. For maximum privacy assurance, verify the current privacy policy and disable optional online features if privacy is critical.
Can I use Loci without Wi-Fi?
Yes, for chat. Once the app and a model are installed, on-device inference works without any internet connection (airplane mode is fine). However, web search, Windows voice input, model downloads, app updates, and any cloud-connected features require internet. If you enable DuckDuckGo web search and use it, that feature will need connectivity.
Does Loci work on older phones?
iOS: Loci requires iOS 18.0+, which excludes iPhone XS and older. Android: Loci works on most modern Android phones (exact minimum OS version not specified by Loci AI), but performance depends on available RAM and the selected model. Mac: requires a recent macOS version supporting the system foundation model. Windows: generally works on modern Windows 10/11 machines with sufficient disk space for a model.
Can I import my own models (GGUF files) into Loci?
No. Loci limits you to its curated library of ~10 models (Gemma, Qwen, Llama, Phi, etc.). If you want to import custom GGUF files from Hugging Face or elsewhere, Private LLM or PocketPal AI are better choices.
What is the difference between Loci and Private LLM?
Loci: free, 5 platforms (iPhone/iPad/Android/Mac/Windows), curated ~10-model library, minimal setup. Private LLM: Apple-only (iPhone/iPad/Mac), one-time purchase, 140+ models, more quantization/flexibility, more configuration. Private LLM is for users who want maximum model control on Apple devices; Loci is for users who want simplicity across platforms.
Can Loci replace ChatGPT or Claude?
For specific tasks, yes β drafting, brainstorming, summarizing local documents, simple Q&A. For complex reasoning, code generation, live web queries, or high-stakes decisions, cloud models (ChatGPT, Claude) are more capable. Loci is best viewed as an offline-capable local alternative to cloud chat for privacy and connectivity reasons, not as a universal replacement.
How much storage does Loci use?
The app itself is small (~100 MB). Model files depend on which you choose: compact models (Phi-4 Mini, Gemma 4 1B, SmolLM) are 1β3 GB; larger models (Llama 3.2 3B, Gemma 4 4B, Qwen 3) are 2β5 GB. If you have multiple models downloaded, total usage can reach 10+ GB. Plan accordingly on devices with limited storage.
Verdict
Loci is most compelling if your priority is low-friction, on-device AI rather than maximum model control. Several features stand out: Gemma 4 E2B/E4B as the strongest reasoning available on mobile; Loci Link (desktop/phone linking to run powerful models from your phone via a connected Mac or PC); global memory across conversations and model switches; first-class thermal management for extended sessions; and exceptional app stability thanks to Loci AI's unique OS memory handling approach. Real-world testing confirms that downloads work reliably, offline chat functions as advertised, and the app experiences significantly fewer crashes than competing local-LLM apps. The web RAG implementation is equally unique: when online, answers ground in live sources; when offline, it falls back cleanly to on-device knowledge without breaking β and you control the behavior in settings. For users who want private offline chat without technical model-management friction and with strong reasoning for a mobile device, Loci excels. For users who want advanced model control and quantization flexibility, Private LLM (Apple) and PocketPal AI offer more depth; for Android users exploring experimental on-device options, Google AI Edge Gallery offers additional model discovery. The honest assessment: Loci succeeds at simplicity, stability, thermal resilience, and cross-platform consistency. It fails only when you need frontier cloud reasoning or deep model control.
Sources
- Loci official site β product overview, platform downloads, latest app version.
- Loci on the App Store β iOS availability, privacy nutrition label, minimum OS requirements, real user reviews.
- Loci on Google Play β Android availability, user ratings.
- Google AI Edge Gallery on GitHub β feature list, supported models, cross-platform availability.
- Google AI Edge Gallery announcement blog post β September 9, 2025 announcement of audio support and Google Play availability.
- Private LLM official site β pricing, privacy claims, Shortcuts integration, platform support.
- Private LLM model library β available models and quantization formats.