ReleaseBytes
databricks Databricks Blog ·

Databricks Unity AI Gateway adds Model Provider Services

blogaigovernancedatabricksengineer
feature announcement

Databricks introduced Model Provider Services (MPS) in Unity AI Gateway, allowing organizations to securely access and govern external AI models like Meta's Muse Spark 1.1. This feature centralizes API key management, enforces access controls via Unity Catalog, and provides end-to-end observability for usage and spend. MPS aims to simplify the adoption of new AI models by consolidating governance and security across different providers.

  • →Model Provider Services for unified AI model governance
  • →Day-one access to new AI models with built-in governance
  • →Centralized access control and credential management
  • →End-to-end observability and cost attribution
  • →Guardrails and rate limiting for AI model requests
Features (5) ›
  • Model Provider Services for unified AI model governance

    Databricks' Unity AI Gateway now supports Model Provider Services (MPS), enabling centralized registration, access control, and governance of external AI models. This eliminates API key sprawl and unifies management through Unity Catalog.

  • Day-one access to new AI models with built-in governance

    Organizations can now adopt newly released models, such as Meta's Muse Spark 1.1, on day one. Unity Catalog's familiar permissions, rate limits, and guardrails are applied to these providers, ensuring secure and governed access.

  • Centralized access control and credential management

    Model Provider Services leverage Unity Catalog for managing access, allowing standard GRANT/REVOKE privileges. Provider API keys are stored encrypted within Unity Catalog, preventing direct exposure to consumers.

  • End-to-end observability and cost attribution

    Every request routed through Unity AI Gateway is automatically tracked for token usage, latency, and cost. Usage data is available in system tables, and inference tables can log full request/response payloads for audit and troubleshooting.

  • Guardrails and rate limiting for AI model requests

    Guardrails and per-service rate limits can be attached to model provider services, enforcing rules for PII, prompt injection, and unsafe content centrally before requests reach the model. This ensures consistent policy application across all users.

Notes (1) ›
  • Muse Spark 1.1 availability via OpenAI-compatible API

    Meta's Muse Spark 1.1 is now available through Databricks MPS by registering it as an OpenAI provider due to its API compatibility. This allows teams to integrate and use the model with standard OpenAI client configurations and specific headers.

Read the original announcement →

https://www.databricks.com/blog/metas-spark-muse-11-now-available-databricks-fully-governed-unity-ai-gateway

Related releases