モデルの提供

Databricks Model Serving is used by data and AI teams that need to deploy and manage models at production scale, including high-QPS, real-time workloads. Condé Nast relies on it for real-time personalization and recommendations across its brands, You.com uses it to deploy models with the reliability and scale needed for real-time applications, GoGuardian serves billions of inferences daily to protect students in high-stakes moments, and Edmunds uses it to securely access and manage multiple SaaS and open models. This makes it suited to teams running both large-scale batch inference pipelines and mission-critical real-time applications.

Model Serving lets you manage custom ML models such as PyFunc, scikit-learn, and LangChain models, foundation models hosted on Databricks such as Llama 3, MPT, and BGE, and foundation models hosted elsewhere such as ChatGPT, Claude 3, Cohere, and Stable Diffusion. All of these are accessible through a single unified interface and API, regardless of whether they're hosted by Databricks or by another model provider on Azure or AWS. This means you can deploy any model type, from pretrained open source models to custom models built on your own data, on both CPUs and GPUs.

Model Serving enables efficient, serverless AI inference on large datasets across all data types and models, without requiring you to manage infrastructure. It integrates with Databricks SQL, Notebooks, and Workflows so you can apply AI at scale directly within your existing pipelines. With AI Functions, you can execute large-scale batch inference instantly, without infrastructure management, while maintaining speed, scalability, and governance.

Model Serving integrates with AI Gateway to meet stringent security and governance requirements across every served model. This lets you enforce proper permissions, monitor model quality, set rate limits, and track lineage, whether models are hosted by Databricks or by another provider. Governance extends to models that are augmented with retrieval-augmented generation (RAG) or fine-tuned on your enterprise data, with automated lookups and monitoring across the entire AI lifecycle.

Model Serving runs as a low-latency API on a highly available serverless service with both CPU and GPU support, priced either pay-per-token on demand with no commitments, or pay-for-provisioned compute for guaranteed throughput. The service scales from zero to meet demand and back down as requirements change, so you only pay for what you use. Databricks handles infrastructure management and maintenance so your team can focus on delivering business value rather than operating servers.

You can try Databricks for free or schedule a demo directly from the product page. From there you can start with one or more pre-deployed models or deploy your own custom or fine-tuned models, choosing pay-per-token or provisioned-throughput pricing based on your workload. The Model Serving documentation and Foundation Model APIs documentation provide step-by-step guidance for deploying and managing models.