Red Hat’s Q3 briefing maps what is production-ready in OpenShift AI 3.5

Red Hat’s Q3 product briefing is useful less as a launch announcement than as a map of what teams can actually use in Red Hat OpenShift AI 3.5. The 56-page deck explicitly labels features as general availability, technical preview or developer preview, then separates them from the OpenShift AI 3.6 roadmap.

What is ready in 3.5

The strongest production-ready cluster is around distributed inference. Red Hat lists llm-d flow control, tiered KV-cache offload to GPU and CPU, controlled deployments with rollback, and inference observability as GA in OpenShift AI 3.5. Azure AKS and CoreWeave CKS are also listed as GA targets for distributed inference, while Amazon EKS remains a technical preview (slides 17 and 52).

The model gateway has gained GA support for external model providers, OpenAI-compatible body-based routing and external OIDC. Multi-tenancy and usage showback remain technical previews, while external metering and token-level cost attribution are developer previews (slide 20). That distinction matters for platform teams deciding whether the gateway is ready to be an internal service boundary or should stay in evaluation.

OpenShift AI 3.5 also makes the Responses API and the main RAG APIs generally available through OGX, formerly Llama Stack. Red Hat says the GA set includes Files, VectorStores and Conversations APIs, with OAuth2, JWKS and attribute-based access control for multi-tenancy (slide 24).

Training and evaluation move forward

For model customization, Red Hat marks GRPO and reinforcement learning with verifiable rewards as GA. Distributed fine-tuning with Ray is also GA, with Training Hub and the CodeFlare SDK supplied in the workflow and images intended to work in disconnected environments (slide 36).

EvalHub’s server, SDK, CLI, agent skills and MCP server are listed as GA, alongside cluster-level and tenant-level deployment modes and Kueue integration (slide 40). Red Hat also calls out GA support for OpenShift AI on hosted control planes running on OpenShift Virtualization, giving each tenant a dedicated control plane while workloads run in VMs on shared GPU infrastructure (slide 49).

What is not GA yet

The same deck keeps several headline features behind preview labels: MCP Gateway and lifecycle management are technical previews; MCP Catalog is a developer preview; OpenShell sandboxing and agent catalog views are developer previews; and AutoRAG remains an advanced technical preview. Red Hat targets OpenShift AI 3.6 for November 2026, including GA goals for MCP management, AutoRAG and AutoML, but the presentation says forward-looking statements are subject to change (slides 2, 25, 27, 34 and 53–54).

For practitioners, the practical reading is straightforward: 3.5 is a production release for core inference, APIs, fine-tuning and evaluation, while much of the agent-management layer should still be treated as an evaluation track. The recorded briefing provides the accompanying product-management walkthrough.