The upgrade broadens provider support, but integrators should audit `apis:` configuration and test custom MCP adapters before rollout.

The GA release becomes Red Hat’s transition path from Migration Toolkit for Containers, adding PVC movement, pipeline generators and a developer-preview agentic migration factory.

The managed service adds 4.22 cluster deployment, port-443 control-plane access in three regions and a final supported release for RHEL VPC workers.

Red Hat’s support table now leaves MTV 2.12 as the only release in full support, narrowing the supported upgrade path for OpenShift Virtualization migration teams.

The release adds supported tenant-scoped ApplicationSets and console tooling, previews rendered-manifest promotion, and carries an upgrade caveat for Redis HA users.

The release supports isolated multi-GPU workloads on NVIDIA DGX B200 and lets Trustee verify Intel TDX quotes without cluster access to Intel PCS.

The release adds GPU-resident weight preloading and broader distributed-serving controls while changing multimodal request handling and several operator-facing flags.

The planned final 3.27 update combines a broad security rollup with an explicit migration signal for teams still on that LTS stream.

The first Quarkus 4 beta moves the baseline to Java 21 and asks extension developers to migrate before a planned late-November final release.

The integration runtime aligns with Camel 4.22.1 and Quarkus 3.40.1, while the project identifies 3.40 as the final feature line on Quarkus 3.

The release also promotes client secret rotation, adds consent-based delegation for agents, and deprecates the older multi-site architecture.

The narrower support phase may require an additional subscription, and Red Hat is giving 4.14 operators one month to prepare.

The release carries the Quarkus 3.39 code line into a year of support and gives teams a documented path from the previous 3.33 LTS.

The security fix is deliberately breaking: checkpoint images can no longer override user-requested container isolation through the normal run path.

The release adds TEE-based model decryption, progressive traffic controls, CPU-backed KV-cache offload and first-class tracing for Kubernetes inference workloads.

The Hermes release gives applications one model name while routing among qualified backends, and adds new evaluation, safety and operational controls.

The release separates signed production images from unsigned pull-request builds while deprecating several llm-d-maintained artifacts.

The Rust gateway adds opt-in telemetry, preserves images across response continuations and broadens its server-side search and MCP handling.

Hermes adds virtual model names, bounded multi-model recipes, open routing models and route-level evaluation for teams operating mixed inference fleets.

The release notes warn that 1.4.0 can trigger authentication, API-key and gateway failures, while 1.4.3 closes policy-targeting and stability defects.

The release can drive host container engines and accelerators from Fedora Toolbox while changing model serving from network-visible to local-only by default.

Direct KEDA configuration, registry-backed model fetching and GPU-aware rollout controls reduce the number of separate mechanisms serving teams must operate.

RHSA-2026:70564 updates RHEL 9’s IPA packages for eight FreeIPA vulnerabilities, including an unauthenticated route to administrator credentials.

The vLLM plugin adds paged KV caching, packed attention and an OpenAI-compatible server for multi-request inference on Macs.

The GA refresh packages vLLM 0.24 and new accelerator stacks, but operators on RHEL AI 1.5 cannot upgrade directly.

The release introduces a persistent GPU weight cache, Gumbel-max watermarking and broad model support while removing several 0.29-era interfaces.

The release prepares meshes for RHEL 10, expands Gateway API support and gives operators a path to mix sidecar and ambient workloads.

The GA release adds default network policies, moves observability to OpenTelemetry and introduces CEL-based PipelineRun placement as a Technology Preview.

The operator’s runtime switch brings newer Camel and Quarkus defaults, but existing deployments need to review compatibility, security context and deprecated platform settings.

The upstream release requires migration work for catalog customizations, search backends, TechDocs plugins and scaffolder templates.

The release moves model checks, inference observability and multi-tenant resource controls into the platform as Red Hat targets production AI operations.

The technical preview exports ingestion and query configurations, adds multilingual model selection and supports pgvector alongside Milvus.

RHEL 9.8 and 10.2 package Python 3.14’s supported free-threaded build alongside the standard interpreter, giving CPU-bound threaded applications a migration path—and a new race-condition test.

Cisco AI POD for Splunk 1.0 turns a Red Hat OpenShift deployment into a validated infrastructure package for Splunk’s AI tools.

The consequential changes sit in route hardening, AI-tool governance and compatibility checks that teams should test before adopting the new upstream LTS.

The upstream release gives administrators opt-in recovery from short GitHub and GitLab failures while restricting where GitHub credentials may be sent.

IBM’s new image option removes the package manager, disables plain HTTP by default and preserves Liberty’s existing application-image workflow.

The Dogtag-based certificate authority adds ML-DSA signatures and ML-KEM key encapsulation while integrating with RHEL cryptographic policy and automated enrollment.

The release makes model evaluation, workload priority, controlled rollouts and observability generally available while keeping several agent and cost controls in preview.

The release changes the default execution path, removes ten deprecated model architectures and ships new admission-control and model-support features.

The release makes plain Camel Quarkus the default, deprecates IntegrationPlatform and adds explicit multi-namespace watching.

Quarkus 3.39.2, 3.33.3.2 LTS and 3.27.5.2 LTS are security-driven maintenance releases; downstream Red Hat users should verify product-specific advisories rather than assume upstream version parity.

The out-of-tree backend preserves vLLM’s OpenAI-compatible surface while adapting scheduling, parallelism and sampling to Tenstorrent mesh hardware.

The stable release also adds agent-ready integration checks, tracing and a commit-SHA guard for fork approvals.

The release adds default-deny network policy, blocks environment variables associated with code injection and makes revision-specific Git clones shallow.

The Important-rated release updates ACM’s multi-cluster data plane and addresses OVN-Kubernetes, Azure, service-discovery and lifecycle failures.

The early-access branch adds new modules and lifecycle changes that OpenShift AI operators should test as an integrated platform, not as isolated components.

The update reaches internet-facing HTTP/2 and authentication paths, but several fixes matter only when specific Tomcat features are enabled.

The new containerized deployment keeps selected host data and recommendation processing inside Satellite, but it replaces hosted console services for managed hosts.

The open-source release adds an end-to-end path for training and serving draft models that predict several future tokens in one pass.