Red Hat today added a range of capabilities, including prompt management tools and an evaluation hub for testing components, to its platform for deploying artificial intelligence (AI) applications.
Unveiled at the Red Hat Summit, Red Hat AI 3.4 also provides access to automated safety testing and red-teaming capabilities based on open source software originally developed by Chatterbox Labs and the Garak open source projects. Red Hat acquired Chatterbox Labs in 2025. The Garak project is advancing a vulnerability scanner for large language models.
Those capabilities are enabled by an instance of open source MLflow software that provides a governance framework for managing the AI artifacts. Red Hat has also added observability and AI agent tracing tools, a Model Context Protocol (MCP) server and gateway and AI agent identity and lifecycle management tools.
A set of Red Hat Desktop tools for building AI applications that now make use of a Podman Desktop tool for building container applications is also generally available. That latter capability enables developers to run AI agents in a sandbox environment. There is also now a dedicated skills repository through which AI agents can invoke Red Hat subscriptions.
Red Hat also announced general availability of support for Red Hat Enterprise Linux (RHEL) running on NVIDIA Grace Blackwell processors with additional support for NVIDIA Vera Rubin processors planned. Red Hat is also making available a developer preview of RHEL on the NVIDIA DGX Spark workstation, in addition to making Red Hat Device Edge available on NVIDIA Jetson Orin developer kits for building AI applications for embedded systems. Red Hat has also joined the OpenShell project launched by NVIDIA to provide a secure runtime for running AI agents.
In addition, Red Hat is also making available an instance of its Fedora distribution of Linux that has been optimized for AI agents. Dubbed Fedora Hummingbird Linux, it allows for anonymous, agent-driven pulls and instant deployment across the hybrid cloud.
In general, Red Hat is advancing a Model-as-a-Service (MaaS) approach based on an AI gateway embedded into the Red Hat AI platform that, in addition to enabling organizations to more easily adopt multiple AI models, enables them to deploy AI applications in the cloud or an on-premises IT environment, says Joe Fernandes, vice president and general manager for the AI Business Unit at Red Hat.
At the core of that effort are a set of vLLM and llm-d frameworks that enable IT organizations to build and deploy distributed AI applications. “We’re trying to enable IT administrators to become model-as-a-service providers,” says Fernandes.
There is, of course, no shortage of platforms for running AI applications. Red Hat is making a case for a platform for running AI inference workloads that provides IT teams more flexibility using a set of curated open source technologies that Red Hat has integrated on their behalf as part of a subscription licensing model.
Regardless of approach, the one thing that is certain is the number of AI inference workloads being deployed in the enterprise is exploding. The challenge now is determining how best to manage those workloads alongside any number of existing monolithic and cloud-native applications that have already been deployed.
