Red Hat AI

Red Hat AI is a hybrid-cloud platform for building, serving and governing enterprise AI workloads on Kubernetes. It bundles Red Hat AI Enterprise, OpenShift AI, RHEL AI and Red Hat AI Inference so platform teams can run any model on any accelerator on-premises, at the edge or across clouds, with pre-deployment safety benchmarking, multi-tenant GPU controls and per-user token metering.

Red Hat AI is Red Hat's umbrella enterprise AI platform, built on OpenShift and positioned around a single claim: any model, any agent, any hardware accelerator, anywhere. It comprises Red Hat AI Enterprise (a unified model and application lifecycle environment), Red Hat OpenShift AI (predictive and generative model lifecycle at scale on Kubernetes), Red Hat Enterprise Linux AI for single-server LLM deployment, and Red Hat AI Inference, which is powered by vLLM. Version 3.5, generally available 9 September 2026, moved the platform's centre of gravity from getting models into production to running them as governed infrastructure. Its headline addition is EvalHub, now GA: an evaluation orchestration service with a versioned REST API that runs multi-framework benchmarks inside isolated, horizontally scalable Kubernetes jobs, letting teams pre-screen custom models, RAG configurations and agents for risk and emit auditable compliance reports. The validated model catalog gained 20-plus entries including Google Gemma 4, NVIDIA Nemotron 3 and Alibaba Qwen, each carrying Garak security scores and PII-exposure and toxicity assessments. On infrastructure, 3.5 added fair-share GPU scheduling, priority-aware serving that protects real-time inference from background jobs, dynamic GPU capacity borrowing, and hosted control planes on OpenShift Virtualization for VM-level tenant isolation on shared hardware. The llm-d distributed inference scheduler went generally available on CoreWeave CKS and Microsoft Azure with Amazon EKS in technology preview — the first time it runs beyond OpenShift. For agents, the Responses API and AutoRAG reached GA with pgvector and multilingual document support, NVIDIA NeMo Guardrails intercept malicious tool calls, and AI Hub ships starter templates for code review, document processing and research. Observability added GPU utilisation and inference-health dashboards, MLflow visual agentic tracing, and per-user token metering for model-as-a-service showback.

Category

Enterprise Platform

Pricing

Subscription, Contact for pricing

Target Market

CIOs, CTOs, Platform Engineers, MLOps Teams, Enterprise Architects

Deployment

Hybrid, Self-hosted, Multi-cloud, Cloud-first

Founded

1993

Headquarters

Raleigh, United States

Team Size

500+

• ✓
  EvalHub safety benchmarking

  Generally available evaluation orchestration service that runs multi-framework benchmarks in isolated Kubernetes jobs and emits auditable compliance reports before deployment.

• ✓
  Validated model catalog

  Over twenty vetted models including Gemma 4, Nemotron 3 and Qwen, each shipping Garak security scores plus PII-exposure and toxicity risk assessments.

• ✓
  llm-d distributed inference

  Distributed inference scheduler now generally available on CoreWeave CKS and Microsoft Azure, with Amazon EKS in technology preview and new KV-cache reuse scorers.

• ✓
  Multi-tenant GPU controls

  Fair-share scheduling, priority-aware serving and dynamic capacity borrowing let several tenants share GPUs without background jobs starving real-time inference.

• ✓
  Per-user token metering and observability

  Dashboards for GPU utilisation, inference health and model performance, plus token metering that makes model-as-a-service chargeback and showback possible.

• ✓
  AutoRAG and agent tooling

  GA Responses API, pgvector-backed multilingual RAG, NeMo Guardrails that intercept malicious tool calls, and starter templates for code review and document workflows.

• ✓
  Hosted control planes on OpenShift Virtualization

  Gives each tenant a dedicated cluster control plane with VM-level isolation while still sharing the underlying accelerator hardware.

• •
  GPU-as-a-service for internal teams

  A platform team sells shared accelerator capacity internally, using fair-share scheduling to prevent contention and token metering to attribute cost per department.

• •
  Regulated model approval

  A risk function runs EvalHub against a fine-tuned model and its RAG pipeline, producing safety and PII-exposure evidence before the model reaches production.

• •
  Sovereign or on-premises inference

  An organisation that cannot send prompts to a public API serves open models in its own datacentre while keeping the same tooling as its cloud footprint.

• •
  Multi-cloud inference standardisation

  A team runs llm-d on CoreWeave and Azure alongside OpenShift on-premises, keeping one scheduler and one operational model across all three environments.

• •
  Governed internal agents

  Developers build document-processing or code-review agents from AI Hub templates while NeMo Guardrails block malicious tool calls and MLflow traces every agent step.

Red Hat publishes subscription guides but no list prices, so every deal is quoted. The metering unit is what matters: self-managed OpenShift AI requires a separate Red Hat AI Accelerator subscription for each discrete physical accelerator card — GPU, TPU, NPU, FPGA or DPU — that is not part of the CPU package, at the same rate regardless of OpenShift edition, so cost scales with GPU count rather than users or tokens. Red Hat AI Enterprise instead provides a unified entitlement covering unlimited accelerators per node, which changes the arithmetic sharply for dense GPU servers. Trials are offered; there is no free tier.

This page was written from 6 sources, 4 on domains other than redhat.com.

1. 1.redhat.com — aivendor
2. 2.redhat.com — red hat puts safety and observability core enterprise ai redvendor
3. 3.techzine.eu — red hat ai 3 5 focuses on governance and shared gpus
4. 4.techstrong.ai — red hat extends scope and reach of ai platform
5. 5.channellife.news — red hat ai 3 5 expands ai safety and observability tools
6. 6.hn.algolia.com — search