GitLab Expanded Support for Open-Source Models for Closed Networks

GitLab 19.0 expanded support for open-source models for self-hosted Duo Agent Platform. This is especially important for teams working in isolated networks who cannot send source code to the cloud.

Air-Gap No Longer Means Falling Behind

Teams working in fully isolated networks without internet access have historically been the last to get access to new AI capabilities. The reason is not technology, but law: in regulated industries, compliance requirements prohibit sending source code to third parties. Previously, GitLab offered limited model choices for such environments, creating a double problem. On one hand, you need a powerful model for complex reasoning—analyzing large diffs, multi-step tools. On the other, simple tasks like renaming a variable don't require computational power. Companies had to either overpay for overkill or work with poor quality on complex tasks.

Four New Models for Different Scenarios

GitLab added support for four open-source models. All have been tested on real Duo Agent Platform tasks: multi-step tool usage, code generation, working with large diffs and multi-file codebases.

• Mistral Devstral 2 123B—focuses on code generation, handles code writing best of all
• GLM-5.1—multilingual model, suitable for international teams
• Kimi-K2.6—distinguished by logging and multi-step reasoning
• MiniMax-M2.7—the lightest, suitable if computational resources are limited

The selection is not random. GitLab tested candidates on the exact tasks the platform solves. Engineers evaluated code generation quality, instruction adherence, and ability to work with large context.

Two Deployment Options

The primary option is your own hardware with vLLM (GitLab's recommended platform for serving open-source models). Computation remains on your server, data never leaves your local network. This is ideal for environments with data residency requirements.

For teams without large capital for hardware, there's an alternative: GPU virtual machines in a private cloud (AWS, Azure, and others). You pay only for what you use, without expenses for purchasing and maintaining equipment. Data isolation guarantees remain the same—nothing goes out to the public internet.

How to Choose Your Path

Choice depends on company requirements. If you need full air-gapped isolation—only self-hosted models on your own infrastructure. If compliance allows a hybrid approach—you can use different models per-feature: for example, simple refactorings on lightweight MiniMax, and complex analysis on Kimi.

What This Means

Air-gap no longer equals falling behind on AI. Regulated industries can now deploy agents of the same quality as corporations that send data to the cloud. Compliance and security should no longer be obstacles to AI productivity.