# Large Language Model Support

> Large language models (LLMs) power the agentic and generative AI features in Automation Suite. The AI Trust Layer governs how these models connect to UiPath products across your organization, applying policy enforcement and audit logging to all AI interactions regardless of which model you use.

Large language models (LLMs) power the agentic and generative AI features in Automation Suite. The AI Trust Layer governs how these models connect to UiPath products across your organization, applying policy enforcement and audit logging to all AI interactions regardless of which model you use.

Automation Suite supports two broad integration modes: connecting to cloud-hosted models through your own subscriptions, and running open source models on your own infrastructure. In both cases, the AI Trust Layer remains the central governance layer. LLM integration is supported on AKS/EKS and OpenShift deployments.

## Integration modes

### Cloud models

Cloud models are LLMs hosted and managed by third-party providers. You can connect your own subscriptions from the following providers:

* **OpenAI** — Available via direct API or Azure-hosted deployment (Microsoft Azure OpenAI).
* **Google Gemini** — Available via Google Vertex AI.
* **Anthropic Claude** — Available via the Amazon Web Services connector.

Cloud model connections give you two configuration options.

#### Bring Your Own Subscription (BYOS)

Replace a UiPath-managed model subscription with your own subscription for the same model family and version. This lets you route model calls through your own account while keeping the UiPath product experience intact.

#### Bring Your Own Model (BYOM)

Add a custom model not included in the UiPath-managed offering, as long as the model passes the AI Trust Layer probe validation test.

UiPath recommends Google Gemini for the broadest compatibility across agentic capabilities.

For configuration steps, see [Configuring LLMs](https://docs.uipath.com/test-cloud/automation-suite/2.2510/admin-guide/managing-ai-trust-layer#configuring-llms).

### Self-hosted models

For organizations that require on-premises model deployments to meet data sovereignty or compliance requirements, Automation Suite supports a set of recommended open-source (OSS) models. Note that self-hosted models are compatible with a subset of agentic capabilities only.

Self-hosted models must expose an API that follows the OpenAI V1 standard. UiPath connects to them through the OpenAI V1 Compliant LLM connector in Integration Service, and each model must pass the AI Trust Layer probe validation test before it can be used.

Running self-hosted models also requires a supported inference engine and appropriate hardware. Supported inference engines include vLLM and SGLang. For hardware requirements, refer to [Kubernetes cluster and nodes](https://docs.uipath.com/automation-suite/automation-suite/2.2510/installation-guide-eks-aks/kubernetes-cluster-and-nodes).

## Provider and capability compatibility

The following table shows which providers are supported for each product and feature, including recommended self-hosted models where applicable. For specific model versions, refer to [Configuring LLMs](https://docs.uipath.com/test-cloud/automation-suite/2.2510/admin-guide/managing-ai-trust-layer#configuring-llms).

  
    
      Product
      Feature
      OpenAI
      Google Gemini
      Anthropic Claude
      Self-hosted OSS
    
  
  
    
      Agents
      Agent execution
      Supported
      Supported
      Supported
      GLM 4.5, GLM 4.6, GLM 4.6 V, GPT-OSS 120B
    
    
      Evaluations
      Supported
      Supported
      Supported
      GLM 4.5, GLM 4.6, GLM 4.6 V, GPT-OSS 120B
    
    
      Simulations
      Supported<sup>1</sup>
      Not supported
      Not supported
      Not supported
    
    
      Autopilot
      Generation
      Not supported
      Supported
      Not supported
      Not supported
    
    
      Chat
      Not supported
      Supported
      Supported
      Not supported
    
    
      Autopilot for Everyone
      Chat
      Supported
      Not supported
      Supported
      Not supported
    
    
      Coded Agents
      Call LLM
      Supported
      Supported
      Supported
      Not supported
    
    
      Context Grounding
      Advanced Extraction
      Supported<sup>2</sup>
      Supported
      Not supported
      Not supported
    
    
      Batch Transform
      Supported
      Supported
      Not supported
      Supported
    
    
      Batch Transform with Web Search
      Not supported
      Supported
      Not supported
      Not supported
    
    
      DeepRAG
      Supported
      Supported
      Not supported
      Not supported
    
    
      Embeddings
      Supported
      Supported
      Not supported
      Supported
    
    
      GenAI Activities
      Build, Test &amp; Deploy
      Supported
      Supported
      Supported
      TBD
    
    
      Healing Agent
      AI-enhanced recovery
      Supported
      Supported
      Not supported
      GLM 4.5, GLM 4.6, Qwen3 235B
    
    
      Popup governance
      Supported
      Supported
      Not supported
      GLM 4.5, GLM 4.6, Qwen3 235B
    
    
      UI Automation
      ScreenPlay
      Supported
      Supported
      Supported<sup>3</sup>
      Not supported
    
    
      Semantic selectors
      Not supported
      Supported
      Not supported
      Not supported
    
    
      Test Manager
      Autopilot
      Supported
      Supported
      Supported
      Not supported
    
  

<sup>1</sup> GPT 4.1 mini only.

<sup>2</sup> Embeddings only. Advanced Extraction and DeepRAG require a cloud model provider.

<sup>3</sup> ScreenPlay only. Semantic Selectors require Google Gemini.

## Governance

All LLM interactions, regardless of provider, pass through the AI Trust Layer. AI Trust Layer policy enforcement via Automation Ops™ and detailed audit logging apply to all model interactions.

Governance policies are specifically designed for UiPath-managed LLMs. If you disable a UiPath-managed model through a policy, that restriction does not extend to your own configured models of the same type.