Mirai

Execute AI models directly on your device for enhanced privacy and speed.

Freemiumtrymirai.comJun 5, 2026

On‑Device AI Models

A curated collection of efficient models (0.3B to 7B parameters) optimized for local execution, reducing total AI costs by an estimated 40 % compared to cloud‑only alternatives.

Zero‑Latency & Full Data Privacy

All inference runs locally on the user’s device, so responses are instant and sensitive data never leaves the device, guaranteeing complete privacy.

Swift Integration

The SDK is built for speed—most developers can complete an integration in minutes without building or maintaining ML infrastructure.

Smart Routing Engine

Automatically chooses between on‑device and (optionally) server‑side paths to balance performance, privacy, and cost, claiming up to 50 % savings.

High‑Performance Inference Engine

Delivers up to 3× performance improvement on Apple platforms, making AI features feel native even on resource‑constrained devices.

Ready‑to‑Use Models & Tools

Pre‑optimized foundation models and helper utilities remove the need for manual quantization, conversion, or benchmarking.

What is Mirai?

Mirai is an on-device AI platform that enables developers to run high-performance AI models directly inside their applications. By processing everything locally, Mirai eliminates cloud latency, strengthens data privacy, and cuts inference costs. It gives product teams a set of lightweight, optimized models (from 0.3B to 7B parameters) and an SDK designed so that a single developer can integrate AI capabilities in minutes—no dedicated machine learning team required. Mirai targets mobile and desktop apps where speed, privacy, and cost-efficiency are critical.

Core Features

On‑Device AI Models: A curated collection of efficient models (0.3B to 7B parameters) optimized for local execution, reducing total AI costs by an estimated 40 % compared to cloud‑only alternatives.
Zero‑Latency & Full Data Privacy: All inference runs locally on the user’s device, so responses are instant and sensitive data never leaves the device, guaranteeing complete privacy.
Swift Integration: The SDK is built for speed—most developers can complete an integration in minutes without building or maintaining ML infrastructure.
Smart Routing Engine: Automatically chooses between on‑device and (optionally) server‑side paths to balance performance, privacy, and cost, claiming up to 50 % savings.
High‑Performance Inference Engine: Delivers up to 3× performance improvement on Apple platforms, making AI features feel native even on resource‑constrained devices.
Ready‑to‑Use Models & Tools: Pre‑optimized foundation models and helper utilities remove the need for manual quantization, conversion, or benchmarking.
Cross‑Platform Support (Apple first, Android soon): Prioritizes iOS, macOS, and visionOS with plans to expand to a broader device ecosystem.

Use Cases & Considerations

Use Cases

Mobile App Developers Adding On‑Device Intelligence: Embed an AI chatbot, summarizer, or content‑generation feature directly inside an iOS app, with no server costs and instant replies—ideal for note‑taking apps, photo editors, or social tools.
Privacy‑Sensitive Products: Build healthcare, legal, or financial applications where client data cannot touch third‑party servers; Mirai ensures all processing stays on the device, helping meet HIPAA, GDPR, or internal compliance requirements.
Startup MVPs with Limited Budgets: A small team can ship an AI‑powered feature without paying per‑call cloud fees; Mirai’s free tier supports up to 10,000 devices, making experimentation cheap during early growth.
Offline‑First Experiences: Travel apps, field‑service tools, or educational software that must work without a stable internet connection can rely entirely on on‑device inference.
Enterprise Teams Extending Existing Apps: Larger organizations can reduce recurring cloud AI spend by shifting routine inference (e.g., classification, text generation) to the edge, reserving expensive remote calls for the hardest problems.
AI Education & Prototyping: Universities and research labs use Mirai to let students run real‑world models on personal devices without complex server setups, accelerating learning and prototyping.

Limitations & Considerations

Platform Lock‑In (Currently): Mirai’s strongest support is on Apple platforms; Android and other operating systems are not yet available, which may limit cross‑platform projects.
Model Selection Relative to Cloud Giants: While the 7B parameter model offers good quality, it cannot match the breadth of cloud‑hosted models like GPT‑4 or Claude for very complex tasks.
Device Hardware Constraints: On‑device inference speed and memory consumption depend heavily on the user’s hardware; older devices may struggle with larger models.
Learning Curve for Customization: Though integration is simple, extending Mirai with custom fine‑tuned models requires deeper ML knowledge that the core SDK abstracts away.
Feature Gaps (Voice, Streaming): Voice processing and certain real‑time streaming capabilities are still on the roadmap, so apps needing these today must look for complementary solutions.
Scaling Costs Beyond Free Tier: For apps with millions of active devices, the per‑device Pro cost can add up, making it important to project expenses and evaluate the Enterprise tier’s volume discounts.

How to use Mirai

Sign Up & Create a Project: Visit trymirai.com, create a free account, and start a new project. Choose your target platform (iOS/macOS for now) and give it a name.
Install the SDK: Add the Mirai SDK to your app via Swift Package Manager or CocoaPods. The documentation provides the exact dependency lines and minimum OS versions.
Select a Model: Browse the model catalog (e.g., a 1.5B parameter model for chat, a smaller one for classification) and download the optimized bundle. Everything happens inside your project dashboard.
Integrate a Few Lines of Code: Using Mirai’s high‑level API, initialize the model, feed it input, and retrieve output. The typical pattern is Mirai.load(model:) then mirai.predict(prompt:)—the examples in the docs show UIKit and SwiftUI snippets.
Test on a Real Device: Run your app on a physical iPhone, iPad, or Mac to experience zero‑latency inference and verify that no network calls are made. Use the analytics dashboard (Pro plan) to monitor usage.
Deploy to the App Store: Once you are satisfied, ship your app. Mirai’s models are self‑contained, so users will download them as part of the app binary or as on‑demand resources, depending on your setup.

Pricing & Plans

Mirai uses a per‑active‑device pricing model with a generous free tier. The free plan supports up to 10,000 active devices, includes an open‑source MIT license, and is intended for individuals and early‑stage teams who can handle self‑serve setup. The Pro plan costs $0.10 per active device per month and adds an analytics dashboard, standard terms, and support for scaling beyond 10,000 devices. For large deployments, the Enterprise tier offers custom pricing, priority Slack support, implementation assistance, custom models, and volume discounts. As pricing details may change, always check the official Mirai website for the most current information.

Platforms

iOS & macOS Native SDK: First‑class Swift packages with support for iPhones, iPads, and Apple Silicon Macs.
visionOS (Apple Vision Pro): Optimized for spatial computing, enabling fully on‑device AI in augmented reality experiences.
Android SDK (Planned): Officially on the roadmap, aimed at expanding the on‑device promise to the broader mobile market.
Web & API Access: Not the primary interface—Mirai is built for embedding, but the routing engine may allow optional server‑side fallback when needed.
Analytics Dashboard (Web): A companion web app for Pro/Enterprise users, showing device usage and model performance.

Tips & Best Practices

Choose the Smallest Model That Meets Quality Needs: Smaller models (0.3B–1B) use less memory and battery; start there and only move up if output quality is insufficient.
Pre‑warm the Model: Load the model early (e.g., at app launch in the background) so the first user interaction is already instant.
Combine On‑Device with Smart Routing: For tasks that need a larger model or real‑time data, let the routing engine escalate to a server‑side call while keeping routine tasks local.
Monitor Performance with the Analytics Dashboard: Pro and Enterprise tiers let you see device‑level metrics; use them to spot models that cause high energy drain or latency spikes.
Respect Device Resources: Set maximum token counts and timeouts, and avoid running continuous non‑stop inference on battery‑sensitive devices.
Stay Informed on Android & Voice Releases: Mirai’s roadmap includes Android SDK and voice processing; early adopters can often join beta programs and shape the feature set.

Who is Mirai for?

Indie iOS Developers: Solo makers who want to add AI to their app without learning ML or paying cloud bills will find Mirai’s free 10k‑device tier especially attractive.
Mobile‑First Startups: Early‑stage teams can ship AI features quickly, validate ideas, and avoid cloud lock‑in while keeping user data safe.
Privacy‑Conscious Enterprises: Organizations in healthcare, legal, finance, or government that must maintain strict data sovereignty benefit from Mirai’s purely local processing.
Product Managers & Non‑ML Engineers: Teams without deep learning specialists can still launch AI‑powered functionality; Mirai’s API abstracts away model optimization and deployment.
AI Educators & Researchers: Teachers who want students to experiment with language models on their own devices, or researchers exploring on‑device inference limits.
Performance‑Sensitive Creators: Anyone building real‑time assistants, accessibility tools, or creative apps where a 200 ms cloud round‑trip would break the user experience.

Alternatives

View all

Core ML (Apple)

Apple’s built‑in inference framework for running ML models on‑device—free and deeply integrated, but requires you to manage model conversion and optimization yourself.

TensorFlow Lite

Google’s lightweight solution for on‑device inference across Android and iOS; very flexible but demands more ML expertise for setup and tuning.

LlamaEdge (WasmEdge)

Run open‑source LLMs locally on edge devices via WebAssembly; great for cross‑platform portability but less plug‑and‑play for mobile app developers.

Replicate

Cloud‑based model hosting with a per‑inference pricing model; convenient and broad model choice, but adds latency and sends data off‑device.

Hugging Face Inference Endpoints

Easily deploy transformer models as managed APIs; strong community and model variety, but still a cloud‑dependent approach with ongoing costs.

Banana ML (or similar serverless GPU providers)

Low‑code GPU‑backed inference; suitable for experiments, but lacks the zero‑latency privacy guarantee of on‑device execution.

FAQ

Q1. What exactly does “on‑device” mean—does no data ever leave the phone?

Yes. All model inference is performed locally on the user’s device using Mirai’s embedded engine. No prompts, outputs, or user data are sent to any external server unless you explicitly configure the smart routing engine to fall back to a cloud provider.

Q2. How does Mirai ensure that the models are optimized for Apple hardware?

Mirai pre‑optimizes models for Apple’s Neural Engine, GPU, and CPU using techniques like quantization and Core ML‑compatible conversion. The inference engine dynamically selects the best available processing unit, often delivering up to 3× speed improvements compared to generic on‑device runtimes.

Q3. Can I fine‑tune or bring my own custom model?

The Enterprise plan includes custom model support, allowing you to upload fine‑tuned versions or models with specific capabilities. For non‑Enterprise users, the current catalog provides a range of ready‑to‑use models, but custom model upload is not yet openly available.

Q4. Is Mirai suitable for a React Native or Flutter app?

Currently, Mirai offers a native Swift SDK. You can bridge it into cross‑platform frameworks via a native module, but there is no first‑class JavaScript/Dart API yet. Many teams wrap the native SDK to expose the functionality they need.

Q5. Does Mirai work offline, and how are model updates handled?

Once the model bundle is downloaded and initialized, all inference works without an internet connection. You control when to push updated models through your app’s update cycle—for example, shipping a new model version in an app update or fetching it on‑demand if network conditions allow.