Red Hat AI 3.4 adds AgentOps and Model-as-a-Service for hybrid cloud
Red Hat has released version 3.4 of its Red Hat AI platform, introducing a set of capabilities designed to move enterprise AI deployments from isolated pilots to governed, production-scale operations across hybrid cloud environments. The release is expected to be generally available by the end of June 2026.
The headline addition is Model-as-a-Service (MaaS), which gives platform engineers a single, governed interface for exposing curated models to developers via OpenAI-compatible API endpoints. MaaS integrates with enterprise identity providers for authentication and consolidates governance across both internally hosted models and external API calls. Red Hat says this addresses a persistent organisational friction: builders want frictionless model access while infrastructure operators need consumption visibility and policy enforcement.
AgentOps and the autonomous agent stack
Red Hat AI 3.4 introduces AgentOps, a set of tools that manage autonomous agents across their full development-to-production lifecycle. The capabilities include integrated tracing via OpenTelemetry, observability over reasoning steps and tool calls, and cryptographic identity management using SPIFFE/SPIRE to replace static API keys with short-lived tokens. The company says this combination allows enterprises to produce a verified audit trail for every agentic action, addressing a risk that has deterred regulated industries from deploying autonomous systems at scale.
The inference layer underpinning the platform is powered by vLLM and the llm-d distributed inference engine. Speculative decoding, now generally available in this release, is reported by the company to improve response speeds by two to three times with minimal quality impact. Request prioritisation allows latency-sensitive interactive traffic to share an endpoint with background workloads. The release also adds day-zero hardware support for NVIDIA Blackwell GPUs and AMD MI325X accelerators, and extends availability to managed cloud environments including CoreWeave, Azure Kubernetes Service, and IBM Cloud.
For safety and red-teaming, Red Hat AI 3.4 integrates Garak, an open-source adversarial scanning tool, alongside technology from Chatterbox Labs and NVIDIA NeMo Guardrails for run-time protection against jailbreaks, prompt injection and bias. MLflow provides experiment tracking and end-to-end tracing of LLM calls, embeddings and retrieval-augmented generation pipelines.
Joe Fernandes, vice president and general manager of the AI Business Unit at Red Hat, said the company is "defining the open standard for how the enterprise executes AI" by providing what he described as a hardened, metal-to-agent foundation for inference, MaaS and AgentOps.
Market context and competitive positioning
Red Hat's framing of an enterprise "token provider" model reflects a broader market shift. As agentic workloads multiply inference demand, organisations are increasingly reluctant to route all traffic through hyperscaler inference endpoints, citing cost unpredictability, data-residency obligations and latency requirements. Red Hat competes in this space against its own parent company IBM, which also offers watsonx infrastructure tooling, as well as purpose-built AI platforms from Databricks, DataRobot and a growing cluster of MLOps vendors.
The open-source underpinning of Red Hat AI 3.4, notably its use of vLLM, MLflow, Garak and SPIFFE/SPIRE, aligns with the digital-sovereignty priorities of European enterprise buyers who are wary of proprietary lock-in. The EU AI Act's obligations for general-purpose AI systems and high-risk deployments are phasing in through 2026 and 2027; audit-trail and transparency features such as those introduced in AgentOps are directly relevant to compliance with those requirements. Similarly, the SPIFFE/SPIRE identity layer maps onto zero-trust architecture mandates increasingly adopted in financial services and critical national infrastructure.
Red Hat has not disclosed pricing for MaaS or AgentOps tooling, nor has it named any early-access customer deployments for the 3.4 feature set. Analysts and buyers will be watching for concrete benchmark data and named production references as the platform enters general availability.