News

    View all →
    Ai-nyan
    Close

    AI ×
    Embedded

    What ailia Enables

    Taking AI all the way to the device.

    ailia supports the whole path — selecting, training and distilling AI models, quantizing them, optimizing for CPU / GPU / NPU, and implementing them in your application.
    Because we hold both AI and embedded engineering in house, we design for accuracy together with speed, memory, power consumption and the constraints of your OS and SoC.

    If no existing model meets your requirements, we build the model itself. We connect to whichever backend suits the target, through ailia SDK, the AI framework we develop ourselves.
    ailia is not locked to a particular model or device. We take the AI you need all the way to something usable in a real product.

    • 420+
      Verified pre-trained models
      Browse ailia MODELS
    • 5OS
      Windows / macOS / Linux / iOS / Android
    • 2017-
      Building our own AI framework, shipping in mass-production products since 2019
      See ailia SDK
    A Venn Diagram Illustrating Ailia at the Intersection of AI and Embedded Technology
    Integrating AI and Embedded Technology into a Unified Process
    AI implementation on embedded devices in detail　｜　Embedded AI

    Three capabilities

    MODEL ENGINEERING
    Build the model
    Training and distillation, LLM / Speech / Voice, custom models for your use case.
    AI COMPUTING
    Run it on the device
    CPU / GPU / NPU, OS / SoC, runtime, on-device optimization.
    CO-DESIGN
    Move between the two
    Iterating between model design and device implementation to meet product requirements.

    Special Movie

    Accelerating AI adoption in society.

    PLAY

    Model
    Engineering

    From Model Engineering
    to AI Computing

    The power to build AI, and the power to run it.
    Both in a single process.

    Model Engineering is the power to build the model your requirements call for — not only selecting an existing model, but distillation, fine-tuning, training from scratch and evaluation. If nothing fits, we train it ourselves. We also offer models tuned specifically for the ailia SDK.

    AI Computing is the power to make AI run efficiently on real hardware and in real products. Drawing on what we have learned building the ailia SDK in house, we cover quantization, model conversion, backend selection, CPU / GPU / NPU optimization, OS / SoC support, pre- and post-processing, and application integration.

    A diagram illustrating the process from customer needs through Model Engineering, Model Assets, and AI Computing to the final product.Tap to enlarge
    Model Engineering and AI Computing — The Power to Build AI and the Power to Run It

    From model selection to mass production

    1. CHOOSE
      Select the best existing model
      Comparing our verified assets and the latest OSS against use case, accuracy, latency, memory and license terms.
    2. BUILD
      If none fits, train or distill one
      Fine-tuning, distillation and custom model development. Model and runtime designed together.
    3. OPTIMIZE
      Fit it to the hardware
      INT8 / INT16 / INT4 quantization, model conversion, operator fusion, memory optimization.
    4. DEPLOY
      Run it on the right backend
      Built around ailia SDK, bringing in QNN or llama.cpp where the device and the model call for it.
    5. OPERATE
      Carry it through to mass production
      On-device evaluation on Windows / macOS / Linux / iOS / Android and embedded targets, through to product implementation and updates.
    Model optimization and device implementation in detail　｜　AI Computing

    Developer products

    What we have built through Model Engineering and AI Computing, packaged so you can use it directly in your implementation.
    ailia SDK sits at the core, with the LLM, Speech and Voice libraries around it.

    Try it first

    Experience AI in the app
    Object detection, pose estimation, super-resolution, speech recognition, text-to-speech, LLM and more — the AI that ailia SDK runs, in an app you can try right now.
    ailia AI Showcase
    • Google Play
    • Mac App Store
    Use the ailia SDK trial
    Try AI demos on the web, and download a one-month trial to evaluate on your own hardware.
    Start Free Trial

    Examples

    MODEL ENGINEERING
    When existing models fall short, we build our own
    The in-house distilled model behind ailia AI Voice, and our own trained d-vector added to DFN3 for speech noise removal — model assets we build ourselves to meet the requirement.
    See ailia AI Voice
    AI COMPUTING
    Making models work in real products
    Because we build the ailia SDK ourselves, we can connect to NPUs through QNN and others. Framework optimization, quantization, OS / SoC support and application integration, to reach the performance the device needs.
    See AI Computing
    Case
    In production
    Read more
    Close

    Technology
    & Engagement

    Publishing what we learn, and talking it through

    We publish what we learn.

    How to choose a model, distillation and quantization, optimization for QNN and NPUs, implementing LLM and speech AI.
    We publish what we have learned on our Technical Blog and in our documentation.

    AI COMPUTING
    Embedded AI, QNN, NPU, Vulkan / Metal, OS / SoC optimization
    MODEL ENGINEERING
    Distillation, quantization, custom training, speech noise removal
    LLM / SPEECH / VOICE
    LLM, Whisper, TTS, agents, backend abstraction
    PRODUCTION KNOW-HOW
    On-device benchmarks, production case studies, what we learned implementing and optimizing

    Technical Blog

    Implementation-level articles on distillation, quantization and backend optimization.

    Documentation

    APIs, sample code and implementation notes for each supplemental library.

    See Documentation

    From concept and model design to devices and mass production.

    "Which model should we choose?" "Will this SoC give us the performance we need?" "Existing models are not accurate enough."
    From questions like these through training and distillation, quantization, on-device optimization, application implementation and mass production — talk to us.

    Model optimization and on-device implementation in detail　｜　AI Computing

    Philosophy

    Philosophy
    AI frees people.
    Throughout our long history,
    we have created new technologies every day,
    building society by harnessing those technologies
    to shape a better tomorrow.

    AI, too, was created by human hands
    and has walked alongside us.
    AI is now indispensable for everyday life —
    it has already woven itself into our world.

    But a good model alone does not make a product.
    AI becomes useful only when it runs on a device,
    fast enough and small enough.
    With the technology to build models
    and the technology to deliver them to real-world devices,
    ailia connects that final stretch.

    AI frees people.
    We aspire to a more free and prosperous future.
    It is already within reach.
    ailia SDK corporate list
    ※Some content on this website uses AI.
    Close
    ailia brand official character
    Ai-nyan
    A mysterious creature from the future who knows all about AI — but doesn't seem to be a cat. Provides information about ailia brand products and services. [Favorite food] Cat food.