Use an API or run locally

    Choose how to serve the model, then compare hardware and engineering options.

    Choose an API

    Connect AI to an existing workflow or your own app. Choose the model and the service that hosts it separately.

    One provider, a first integration

    Start with the direct OpenAI or Claude API for the model you already tested. Compare a small set of your own tasks before committing to a provider.

    Existing cloud and company controls

    Use Azure, AWS or Google Cloud when procurement, identity or regional hosting requirements lead the choice. Check the exact model version and supported features on that service.

    Several model providers

    An aggregator such as OpenRouter can simplify switching. Compare its routing, data handling, failure recovery and bill against a direct API.

    Test success rate, latency and total cost on the same tasks. Include tools, retries and review time. App subscriptions and API billing are separate; keep API keys on the server and set a spending limit.

    Provider endpoints & official docs →

    Run AI on your own computer

    Choose a local runtime, fit the model to your memory, then test your real task before buying hardware.

    Setup references checked 2026-10-22

    A desktop interface: LM Studio

    Download a supported model, load it, then chat or work with documents. Once the model and runtime are downloaded, core chat, document processing and the local server can work offline.

    LM Studio · offline operation ↗

    A local endpoint: Ollama

    Install Ollama and choose a local model. Its documented starter example uses Gemma 4 E2B; confirm available memory and model requirements before downloading. Cloud models follow a different data path.

    ollama pull gemma4:e2b
    Ollama · local API quickstart ↗
    1. Start with a small quantized model that fits available memory. Weights, context cache and runtime overhead all consume memory; a model file fitting on disk is not enough.
    2. Try a real Korean document or coding task. Check correctness, time to first output, generation speed and the amount of correction needed.
    3. Connect the local endpoint to a compatible harness only after testing chat and tool calling. Downloaded models, external tools and remote APIs have separate privacy and licensing implications.
    Compare GPUs, Mac Studio & NPUs →Open-weight model families →

    Use hosted AI when setup and maintenance cost more than the problem requires. Local inference is useful when offline access, data control or sustained utilization justify its hardware and operating cost.

    Build & deploy

    Turn a useful workflow into something repeatable. Choose the next guide for the job you are doing.

    Agent skills →

    Package a repeatable task and connect the tools it needs.

    AI engineering →

    Build, evaluate and operate an AI system.

    Model APIs →

    Find provider endpoints and integration references.

    Inference hardware →

    Compare server costs and throughput before choosing GPUs, a Mac or NPUs.

    Core guide checked · Individual sources carry their own dates.

    Report an outdated instruction