Enterprise ·

Your Models, Anywhere

Why enterprise AI needs to connect to models wherever they live—cloud, on-premises, or air-gapped.

The enterprise reality

Most AI interfaces assume you'll use their model, through their API, with your data leaving your network. That works fine for consumer applications. It doesn't work for many enterprises.

Banks can't send customer data to external APIs. Healthcare systems have HIPAA constraints. Defense contractors operate in classified environments. Many organizations have legitimate reasons to run AI on their own infrastructure.

At the same time, organizations don't want to be locked into a single model provider. They want the flexibility to use the best model for each task—whether that's Claude for analysis, GPT for generation, or a local model for sensitive queries.

The interface problem

Every model has a different interface. Cloud APIs work one way. Local models work another. Your organization's private fine-tuned model works yet another way.

Users shouldn't need to care about this. They have questions; they want answers. The plumbing should be invisible.

What users actually want

  • One interface for all models
  • Ability to compare responses across models
  • Control over which models see which data
  • Same experience on web, desktop, and mobile

Where models live

☁️

Cloud APIs

The standard approach. OpenAI, Anthropic, Google, xAI—managed services with simple API access.

Best for: General queries, rapid iteration, when data can leave your network
🏢

Private cloud

Models deployed in your VPC. Still cloud-hosted, but within your security boundary. Azure OpenAI, AWS Bedrock, GCP Vertex.

Best for: Regulated industries with cloud-first mandates
🖥️

On-premises

Models running in your data center. Full control over hardware, data, and access. Ollama, vLLM, TGI.

Best for: Sensitive data that cannot leave premises
🔒

Air-gapped

Completely isolated networks with no internet connectivity. Models installed locally, no external communication.

Best for: Classified environments, critical infrastructure
💻

Desktop local

Models running directly on user workstations. Apple Silicon, NVIDIA GPUs, or even CPU-only for smaller models.

Best for: Power users, offline work, personal data
🎯

Custom fine-tuned

Your own models trained on proprietary data. Domain-specific, organization-specific, deployed wherever makes sense.

Best for: Specialized tasks, competitive differentiation

The universal connector approach

Rather than building different interfaces for different model sources, we built a universal connector that abstracts the differences:

One Interface
Web • Desktop • iOS • Android
Universal Connector
API normalization • Auth management • Response synthesis
Cloud APIs
OpenAI, Anthropic, Google
On-Prem
Ollama, vLLM, Custom
Air-Gapped
Local only

From the user's perspective, they just pick a model and ask a question. The connector handles authentication, API differences, error handling, and response formatting.

Real deployment scenarios

Bank compliance team

Uses cloud models for general research, on-premises fine-tuned models for regulatory analysis, with automatic routing based on query sensitivity.

Claude (cloud) Llama 8B (on-prem) OSFI specialist (custom)

Healthcare analytics

All patient-related queries stay on-premises. Administrative queries can use cloud models for better performance.

GPT-4 (admin only) Mistral 7B (clinical)

Defense contractor

Entirely air-gapped deployment. Desktop app with local models only. No network connectivity required.

Llama 70B (local) Phi-3 (local fallback)

Research team

Multi-model deliberation. Query five models simultaneously, synthesize results. Best of all worlds.

Claude GPT-4 Gemini Grok DeepSeek

Same experience everywhere

Enterprise users work across devices. The interface should follow them:

Web

Zero install, works anywhere with a browser. Cloud models only.

Desktop (macOS, Windows, Linux)

Native apps with local model support. Global hotkeys, menu bar access, works offline.

iOS

Mobile access to cloud models. Quick queries on the go.

Android

Same mobile experience. Cloud models with secure authentication.

Security considerations

API keys stay local

Your API credentials are stored locally on your device. They're never sent to us. We're just an interface, not a proxy.

Data routing control

Configure which models can see which types of queries. Sensitive data can be restricted to on-premises models only.

Audit logging

Full audit trail of which users queried which models with what. Exportable for compliance purposes.

SSO integration

SAML/OIDC support for enterprise identity providers. Okta, Azure AD, whatever you use.

The takeaway

Enterprise AI isn't about picking one model or one deployment pattern. Different use cases need different solutions. The interface should handle that complexity invisibly.

Your models, wherever they live. Your data, under your control. One interface to access it all.

Deploy multi-AI anywhere

Onyx Legion supports cloud, on-prem, and air-gapped deployment.

Legion Enterprise →