Every Collibra AI feature has, until now, come from the same place: Collibra's own managed models. For an enterprise that already runs its own vetted, compliant AI infrastructure, that’s a second security review nobody asked for. Bring your own model (BYOM) removes the coupling: the AI feature stays, the model underneath it is now yours to choose and control.

What’s new: Bring your own model (BYOM)

BYOM lets a Collibra customer register their own AI model provider and have Collibra's AI features call that model instead of Collibra's managed one. It’s out now in public preview, starting with Google Gemini Enterprise.

Setup happens through a per-provider configuration screen: project ID, endpoint and either a service account or an API key for Google Gemini Enterprise specifically, since the required credentials differ by provider. Environments are restricted to exactly one active model per feature group at launch. The same model can be used multiple times. Behind the scenes, Collibra's AI features don't call a raw model name, they call a tier (collibra/low, collibra/medium, collibra/high), and BYOM changes what each tier resolves to, without changing how the features themselves are built.

Future supported models include anything OpenAI API-compatible: OpenAI’s GPT family, Anthropic’s Claude models, AWS Bedrock, popular openweight models like Qwen and GLM, your own custom models and more.

How BYOM helps

Every enterprise that adopted Collibra's AI features (now under the Maestro family of capabilities) accepted a package deal: the feature, and the model (Google Gemini Enterprise) running underneath it. A compliance team that had already cleared its own AI infrastructure still had to review Collibra's managed model, or block the feature outright. A data science team got Collibra's managed model regardless of what they'd already tuned or licensed elsewhere. BYOM takes the deal apart and provides enterprise greater choice.

Problems it solves

A second vendor review for the same use case: compliance teams had already approved their own model provider, yet still needed a parallel review of Collibra's managed inference path before turning on an AI feature.
Tuned or licensed models going unused: data scientists who'd already fine-tuned or licensed a model outside Collibra got Collibra's managed model regardless.
No say in where inference runs: IT administrators had no way to specify the provider or project their AI workloads ran against.
Provider sprawl across the AI stack: organizations standardizing on one AI vendor still ran Collibra's managed model as a second, parallel provider relationship.
Cost control: Many organizations have spend commits with their AI provider; BYOM allows them to go against that commit using Collibra AI features.

How BYOM works

In Collibra's AI settings, under model providers, select the option for Google Gemini Enterprise. You add a provider, and the form asks for exactly what Google needs: project ID, region, and either a service-account key or an API key. Collibra validates the credentials before you save, then runs a connectivity test against the endpoint.

The Model providers screen showing a Google Gemini Enterprise provider added, credentials validated, and marked as the active provider.

Google Gemini Enterprise becomes the environment's active provider. Within Google Gemini Enterprise, we allow for different (or the same) models to be selected for Low, Medium, and High effort. In the future, we’ll offer the ability to connect multiple providers simultaneously so you can select the one that matches your cost and compliance objectives at that time on a granular level.

Before BYOM, when one of Collibra’s AI features needed a mid-weight model call, it hit the managed collibra/medium tier, backed by a Collibra-hosted model. After you register Google Gemini Enterprise, that same call still asks for collibra/medium (the feature code doesn't change) but the tier now resolves to your Google Gemini Enterprise project's model instead.

The tier-mapping view showing collibra/low, collibra/medium and collibra/high resolving to the customer's active Google Gemini Enterprise models.

What doesn't move is the governance layer underneath. BYOM changes where inference happens; it doesn't change what Collibra already tracks about the metadata behind the request. Registering a provider is a compute swap, not a governance one.

Why should you be excited?

Enterprise IT Administrators: register and monitor a model connection without opening a new vendor relationship or a fresh security review for AI compute.
Data Scientists / ML Engineers: point Collibra's AI features at a model you've already tuned, licensed or validated, instead of a fixed managed model.
Compliance Officers: keep AI processing inside infrastructure that already passed review, closing the objection that had been blocking AI feature adoption.
Collibra Administrators: manage the active provider from one screen, where a single switch governs every AI feature in the environment.

Use cases

Reusing an AI investment already in place: an organization running its approved model deployment through its own Google Gemini Enterprise project points to Collibra's AI features at that same project, instead of standing up a parallel Collibra-managed model for the same workload.
Clearing a compliance blocker: a compliance team that already cleared its Google Gemini Enterprise environment for sensitive data can turn on Collibra's AI features without a separate review cycle for Collibra's managed inference path.
Swapping providers without redoing every feature: an admin who registered Google Gemini Enterprise at launch later moves the organization's active provider as AI vendor strategy changes, and every AI feature follows the switch without individual reconfiguration.

Key takeaways about BYOM

Bring your own model doesn't change what an enterprise's AI features do, it changes who controls the model answering them. Starting now, that control point is a single registered provider per environment, beginning with your instance of Google Gemini Enterprise. What Collibra reviews is the connection; what a model actually produces is a separate, narrower claim, and the two shouldn't be confused when evaluating a provider.

Inference moves, governance doesn't: BYOM changes where a model runs; it does not change what Collibra tracks about the metadata behind the request.
Purpose-built, not generic: each provider gets its own setup, built for what that specific backend requires, there's no generic "paste an endpoint" option.
The switch is platform-wide: activating a provider replaces the active one for every AI feature in the environment at once, not feature by feature.
Collibra CollibraEnterprise AI Control Plane

Collibra

Enterprise AI Control Plane