Version & Feature Tracker
Tracks Red Hat OpenShift AI's feature lifecycle, deprecations, and upgrade paths release over release — cross-checked against the live Supported Configurations matrix so sellers and customers can see where the release notes and the formal support matrix disagree, not just where they agree.
Release timeline
Overlaid with the official OpenShift AI Self-Managed Life Cycle: each card's phase badge (Full Support / Extended Update Support / End of Life) is computed against today's date, not hand-maintained. EA1 → EA2 → GA milestones shown where a release had dated Early Access drops.
What's new in 3.5
Auto-derived from every feature whose most recent timeline entry lands on 3.5.
Automated tool-calling evaluation-data generation for custom MCP servers GA
Actively explores MCP servers to generate verified question-answer-tool-call triplets for single- and multi-step tool chains, feeding standard evaluation frameworks.
MCP Catalog support-tier labeling TP
Badges distinguish Red Hat Supported, Partner Supported, and Community Supported MCP server entries.
MCP Lifecycle Operator TP
Kubernetes-native operator for deploying/managing MCP servers. Current 3.5 implementation guide: mcplifecycleoperator defaults to Removed in DataScienceCluster and must be explicitly set to Managed; mcpCatalog...
MCP gateway Operator TP
Optional external dependency, installed/upgraded independently of the OpenShift AI Operator. Cluster administrators own its lifecycle; used for governed MCP management and tool-call routing.
OGX-native agentic API surface (Tool Runtime, Responses, Conversations) TP
Mixed GA/TP/DP: Tool Runtime API and model-context-protocol tool-runtime provider are DP; Responses API is GA; Conversations API is TP. See ogx_provider_table for full per-endpoint/provider detail.
OpenCode coding-agent deployment TP
Validated coding-agent pattern showing vLLM/OGX inference backends and MLflow tracing generalize to coding-specific agent workloads.
Validated tool-calling configuration in Model Catalog TP
Gated behind spec.dashboardConfig.toolCalling in OdhDashboardConfig; surfaces validated vLLM tool-calling and reasoning arguments for supported models.
Agent Catalog in AI Hub DP
Gated behind agentsCatalog. Central discovery interface for starter kits built on LangGraph, CrewAI, LlamaIndex, and other frameworks.
CSV export for Model Catalog data DP
Standalone Python 3.10+ CLI that paginates Model Catalog REST API data and emits RFC 4180 CSV including custom properties.
Claude Code agent starter kit DP
Validated direct Anthropic, self-hosted vLLM, or vLLM-through-OGX paths; MLflow tracing; ConfigMap skill/MCP injection; restricted-v2 SCC.
External metering integration for MaaS DP
Separate MaaS-specific BBR plugins for post-inference usage webhook and pre-inference credit checking; invoice/rate calculation remains external.
External metering: per-user token usage and cost DP
Inference Gateway IPP plugin plus PostgreSQL-backed metering service; balance checks, provider token extraction, CloudEvents, admin dashboard, REST API.
Gen AI Studio saved-agent configuration DP
Gated behind agentConfigManagement. Saves model, inference, prompt, RAG, and MCP configuration as reusable project-scoped agent resources.
Hierarchical KV Cache Tiering DP
Configurable cache tiers to increase effective cache size and prefix-cache reuse for concurrent long-context and multi-turn workloads.
Kale JupyterLab extension DP
Converts annotated notebooks into AI Pipelines without KFP SDK code; pre-installed but disabled by default in standard images; requires a DSPA in the namespace.
Latency-aware routing for llm-d DP
Uses per-request TTFT/TPOT targets to place latency-sensitive and throughput-sensitive traffic on suitable pods.
LoRA-aware routing for llm-d DP
Routes requests to pods where the requested adapter is already loaded; falls back to standard routing when no matching pod is available.
MCP Catalog administrative interface DP
Administrator-only YAML add/edit/remove for MCP Catalog sources from Settings, replacing manual ConfigMap editing.
MiDojo adversarial testing engine DP
MITM agent-security testing that intercepts tool communications, injects attacks, and grades utility/security using declarative scenarios and multiple agent protocols.
OGX File Processors API DP
inline::auto/docling/markitdown/pypdf and remote::docling-serve providers; some require dependency-specific custom configuration.
OGX remote::anthropic and remote::gemini providers DP
Provider-specific OGX inference integrations enabled by ANTHROPIC_API_KEY and ENABLE_GEMINI.
OpenClaw agent starter kit DP
Validated Kustomize/installer deployment; self-hosted vLLM through OGX; MLflow/OpenTelemetry tracing; browser RBAC; restricted-v2 SCC.
OpenShell (secure agent sandboxing and policy enforcement) DP
Helm deployment, mTLS, LLM provider setup, isolated sandboxes, controlled network egress. Experimental upstream artifacts; not supported for production.
Text-mode training for multimodal models in Training Hub DP
Text-only fine-tuning path for listed vision-language model families; full multimodal fine-tuning not supported in this release.
View running agent deployments in dashboard DP
Lists OpenShell-managed Sandbox custom resources deployed manually, with per-namespace filtering.
KubeRay (CodeFlare Operator successor) GA
KubeRay remains the successor component. Live source conflict: 3.5 release notes say the operator was upgraded to 1.6.x, while the current Supported Configurations matrix lists KubeRay 1.4.2 (GA x86_64/Arm, TP Power,...
Kubeflow Trainer v2 (replacing Kubeflow Training Operator v1) GA
Remains GA on x86_64/Arm. Training Hub fine-tuning algorithms (SFT, OSFT, LoRA, GRPO) can run on Ray clusters; universal workbench images include RLVR/GRPO dependencies; automated prompt optimization added for...
Kubeflow Training Operator v1 (KFTOv1) Deprecated
Current support matrix lists Kubeflow Training Operator v1 1.9.0 as Deprecated on x86_64/Arm/Power, with a component-level EUS exception: supported only through 3.5 Full Support, EOL when Full Support ends 27 March...
Kueue (embedded component -> Red Hat build of Kueue Operator) GA
New at 3.5 GA: administrators can self-manage ClusterQueues/LocalQueues via a boolean flag on the Kueue component spec, disabled by default, so the Operator skips automatic queue creation for GitOps-managed environments.
Automated Red Teaming (powered by Garak) GA
Reaches GA: OpenAI Responses API endpoint support, multilingual red-teaming via LLM-based translation, multiclass judge safety evaluation, parallelized execution, disconnected/air-gapped support, runnable via EvalHub...
Evaluation Stack -> EvalHub GA
3.5 GA release notes state EvalHub is generally available with a documented API-versioning scheme, breaking-change policy, must-gather tooling, and full Red Hat support (SDK/CLI, EvalCard, per-tenant deployment). The...
AutoML TP
Remains TP. Current matrix: TP on x86_64 and IBM Power, unavailable on Arm and IBM Z. Adds model-insight curves/backtesting, experiment transparency, TP AutoGluon serving runtime; no GA promotion.
AutoRAG TP
Remains TP on x86_64 in the current matrix; unavailable on Arm/Power/Z. Adds CPU-only deployment, multilingual support, structural contextualization, visual experiment-pipeline representation. EA2 release notes claim...
Feature Store (Feast-based) GA
Adds SparkApplication batch materialization engine (batch_engine.type: spark_application, feature transformations on Kubeflow-Spark-Operator-managed Spark clusters; supports enterprise RAG pipelines via batch...
FMS-Guardrails Orchestrator (via TrustyAI) Deprecated
Directly confirmed in the 3.5 release notes Chapter 5, Support Removals: deprecated, remains functional in 3.5, will be removed in a future release (undated). NeMo Guardrails is now the single recommended framework...
NeMo Guardrails GA
Gen AI Studio playground guardrails experience rebuilt on NeMo Guardrails (independently toggled input/output rails). Core NeMo Guardrails remains GA on x86_64 and IBM Power. Release notes announce IBM Z support, but...
NeMo Guardrails integration with MCP Gateway TP
Introduced 3.5 EA2, retained TP in current 3.5 docs. Enforces guardrails on agent tool calls at the MCP Gateway layer; requires the MCP Gateway Operator as an external prerequisite. Distinct from OGX model-response...
MLflow experiment tracking / MLOps integration GA
Scope substantially widened: embedded directly into AI Pipelines, workbenches, the Kubeflow SDK, Kubeflow Trainer, and AI Hub. Training jobs run through a pipeline or SDK are tracked by a dedicated per-project MLflow...
Loki-based showback and user-scoped usage dashboards (MaaS) TP
Expanded: Loki-based structured log pipeline with 30-day object-storage retention for monthly billing, plus read-only, data-isolated per-user usage dashboards.
MaaS deployment/routing path for Distributed Inference with llm-d TP
Unchanged Technology Preview. Its status does not change the GA status of core llm-d or core MaaS.
MaaS multi-tenancy with per-tenant gateway and identity isolation TP
New in 3.5 GA. A single CR provisions a tenant namespace, per-tenant API service, and auth/rate-limit policy; gateway and identity provider are external prerequisites. Supports long-lived tenant-scoped API keys,...
MaaS-specific external OIDC user authentication GA
TP -> GA in 3.5. Administrators integrate an external OIDC identity provider, map OIDC group claims to MaaS subscriptions, and issue API keys scoped to OIDC group memberships -- without requiring OpenShift accounts.
Models-as-a-Service (MaaS) core: access model and token mechanism GA
GA core expanded: external OIDC authentication for MaaS reaches GA (map OIDC group claims to subscriptions, issue API keys without requiring OpenShift accounts, coexists with OpenShift-group auth). OpenAI-compatible...
Multi-provider API passthrough for external models (MaaS) TP
Expanded: native-provider passthrough through the MaaS gateway using Anthropic Messages API (/v1/messages) or OpenAI Responses API (/v1/responses), without format translation, preserving provider-specific features...
Platform-wide direct OIDC authentication (GatewayConfig CR) GA
No maturity change; unchanged platform capability. (Distinct sub-feature: MaaS-specific external OIDC user authentication moved TP->GA in 3.5 -- see maas-external-oidc.)
vLLM model deployment directly through MaaS TP
Remains Technology Preview. The current OdhDashboardConfig documentation explicitly marks vLLMDeploymentOnMaaS as Technology Preview; flag defaults to false.
Model registry and model catalog GA
Catalog enrichment: Safety and Security Insights tab (security scan results by category -- prompt injection, jailbreak resistance, harmful content generation). Cold-start load time and minimum vRAM metrics per GPU...
Distributed Inference with llm-d (core) GA
GA extended to Azure Kubernetes Service (AKS) and CoreWeave Kubernetes Service (CKS); Amazon EKS added as Technology Preview (inference-aware scheduling also TP on EKS). Istio is now the supported gateway...
Distributed Inference with llm-d: priority-based flow control GA
Breaking change on promotion: TP (3.4) -> GA (3.5) while simultaneously changing its API group (inference.networking.x-k8s.io -> llm-d.ai), metrics prefix (inference_extension_ -> llm_d_epp_), and saturation-detector...
KServe Serverless deployment mode Removed
Retired from the supported 3.x serving architecture. All Serverless workloads must be migrated to KServe RawDeployment before moving to 3.x. New at 3.5 GA: canary rollout support for KServe RawDeployment...
LLMInferenceService / LLMInferenceServiceConfig (llm-d-native) GA
Maturity follows llm-d: GA on OpenShift/AKS/CKS, TP on EKS as of 3.5.
Multi-model serving platform (ModelMesh) Removed
Retired from the supported 3.x serving architecture. Persistence of legacy objects is not equivalent to product support.
Red Hat AI Inference fast-release images as custom ServingRuntime GA
New in 3.5: deploy Red Hat AI Inference fast release container images as custom serving runtimes on an existing OpenShift AI installation without upgrading OpenShift AI -- access to newer vLLM versions/model support...
vLLM ServingRuntime for KServe GA
Remains GA. Foundational vLLM GA status does not extend to every vLLM-adjacent delivery path -- see the other vllm-* feature entries.
Llama Stack / OGX (Open GenAI Stack) core: RAG, agentic workflows, vector DB integration GA
Renamed to OGX at 3.5 EA1 (breaking manual migration: remove Llama Stack Operator, create new OGXServer resources; API group llamastack.io->ogx.io; v1alpha1->v1beta1; kind LlamaStackDistribution->OGXServer;...
SQLite as Llama Stack / OGX metadata store Deprecated
Unchanged through 3.5 GA: SQLite remains available only for local dev/testing and must be explicitly configured; PostgreSQL required for production-grade OGX deployments.
Gateway API and Red Hat Connectivity Link (RHCL) GA
New known issue on RHCL 1.4.0/1.4.1: rate limiting silently stops working on batch gateway deployments (RHOAIENG-76586), plus an OOMKill risk from the 1.4 wasm plugin's higher compile-time memory requirement --...
⚠️ Breaking changes / high-risk items in 3.5
FMS-Guardrails Orchestrator (via TrustyAI) Deprecated
Directly confirmed in the 3.5 release notes Chapter 5, Support Removals: deprecated, remains functional in 3.5, will be removed in a future release (undated). NeMo Guardrails is now the single recommended framework for all LLM safety guardrailing. The...
Kubeflow Training Operator v1 (KFTOv1) Deprecated
Current support matrix lists Kubeflow Training Operator v1 1.9.0 as Deprecated on x86_64/Arm/Power, with a component-level EUS exception: supported only through 3.5 Full Support, EOL when Full Support ends 27 March 2027. Migrate to Kubeflow Trainer v2...
KServe Serverless deployment mode Removed
Retired from the supported 3.x serving architecture. All Serverless workloads must be migrated to KServe RawDeployment before moving to 3.x. New at 3.5 GA: canary rollout support for KServe RawDeployment InferenceServices ships GA (progressive traffic...
Multi-model serving platform (ModelMesh) Removed
Retired from the supported 3.x serving architecture. Persistence of legacy objects is not equivalent to product support.
Explore
Feature Lifecycle Matrix
Every tracked feature/component across every version, color-coded.
Upgrade Path / Migration Checklist
Every deprecation, removal, and breaking change in one actionable list.
Source Conflicts
Where the release notes and the Supported Configurations matrix disagree.
3.5 release detail
Full chapter breakdown parsed from the official release notes.