Red Hat’s A100 tests show why a valid Kubernetes device allocation can still be a bad inference allocation—and how cross-driver topology matching can prevent it.
Red Hat’s UK CIO survey finds near-universal confidence in agent security despite limited data-flow, dependency and vendor-risk visibility.
A proof of concept uses Asago-generated risk scenarios, a permission recommender and repeat evaluation to narrow what retrieval agents can access.
Red Hat’s new integration pattern exposes existing Camel routes to models while keeping tool visibility, arguments and iteration counts under application control.
A public-sector proof of value used synthetic data, a custom rubric and three L40 GPUs to test specialized models against GPT-4.1.
A new engineering guide asks platform teams to test vendor claims against multi-tenancy, GitOps, failover and storage requirements before deployment.
The Korean payments provider says an OpenShift AI platform using vLLM compression, FP8 quantization and a multi-model design cut GPU inference costs by as much as 70%.
A production Quarkus architecture turns Oracle changes into targeted WebSocket and mobile notifications without polling the system of record.
A new architecture guide argues that catalogs, lineage events, policy translation and semantic safeguards must be designed as one agent-data system.
torch-spyre, a vLLM plug-in and hardware-agnostic model layers give IBM’s inference accelerator a path into the open serving stack.
Ubicloud measured malformed Kimi tool calls on an unconstrained deployment, then reached 100% schema accuracy by enabling vLLM’s new strictness flag.
Red Hat’s Developer Preview replaces the connected advisor pipeline with a single deployment on an Advanced Cluster Management hub.
The field build compiles behavioral memory into a versioned profile, but its stale-state and silent conflict handling make governance as important as token savings.
A GB300 NVL72 study shows that expert placement and host-side scheduling matter as much as kernels when decode steps become short.
A new IBM lab blueprint joins DCGM workload signals, cloud cost data and alerts-as-code so platform teams can separate useful accelerator work from paid-for idle capacity.
Red Hat and vLLM engineers combine token-parallel prefill, sharded KV cache and host-memory decode to fit contexts that standard parallelism cannot hold.
MiniMax-M3.1-Flash arrives with sparse-attention, MoE and speculative-decoding paths for NVIDIA and AMD accelerators, plus reproducible recipes.
An OpenShift pattern combines gateway isolation, network policy, identity-aware tool permissions and rate limits instead of treating authentication as sufficient.
RHEL CoreOS 10 is a Technology Preview in OpenShift 4.22; Red Hat’s planned mixed-stream model would let administrators validate and move nodes gradually.
The clinical-simulation project joins OpenShift orchestration, IBM Fusion storage and NVIDIA inference in a permission-aware, on-premises architecture.
The release gives Kubernetes operators explicit control over pod metadata, adds standards-based OCSF export and hardens failure paths around sandbox policy enforcement.
The open-source plug-in captures and edits model internals during CUDA Graph replay, avoiding the eager-mode penalty that keeps many safety tools out of production.
Red Hat’s new benchmark replaces Python-heavy toy problems with reproducible Go, Python and C tasks drawn from Podman, etcd, Ansible, KubeVirt and related projects.
A production case study shows how queue-aware and prefix-cache-aware routing can turn idle accelerator capacity into predictable training-data output.
The new open-source library separates quantization descriptions from GPU execution choices, then compiles a tuned kernel for each workload.
A first-party reference design joins simulation, evaluation, signed OCI artifacts and fleet delivery on one AI workstation.
A troubleshooting-assistant design replaces flat fan-out with layered orchestrators, trading additional model calls for narrower routing decisions and contained failures.
Red Hat’s two-site OpenShift lab preserved OIDC sessions during site failure, but the Keycloak 26.7 stateless work remains a preview that operators must validate.
IBM’s third-quarter disclosure points to continued momentum in its hybrid-cloud business, but the segment reporting stops well short of showing which Red Hat products drove it.
Day-zero support for Recursive’s sparse model shows how model-specific kernels can turn long-context serving into a single-Blackwell deployment.
The automaker says Ansible cut one large configuration workflow from 60 hours to 20 minutes while replacing a 30-year-old management system.
Red Hat’s engineering guide uses typed probabilities for messy text while keeping thresholds, authorizations and execution in ordinary application policy.
Linux Foundation research sponsored in part by Red Hat pairs reported developer time savings with rising review load, weak contribution policies and burnout risk.
The airline’s 1,200-developer cloud workspace platform moves agent compute off laptops while keeping tools, secrets and nested containers inside governed project environments.
A new llm-d flow-control path prioritizes workloads only when a model-serving pool saturates, with per-tenant fairness and explicit limits for long outputs.
A merged chunk/noisy-KV path brings stage-pipelined autoregressive video diffusion into the upstream serving stack, but it remains explicitly experimental.
NIXL EP and reusable CUDA graphs cut measured topology-change pauses from tens of seconds to less than two, but orchestration integration is still in progress.
The merged path adds OpenAI-style image generation, parallel execution and layerwise offload, but its published performance evidence covers a narrower configuration.
An AWS-contributed connector overlaps destination setup with prefill work, trading later routing flexibility for lower latency under load.
An AWS-contributed path lets disaggregated vLLM deployments reuse generated context across turns instead of repeatedly rebuilding it.
Red Hat’s latest engineering update shows the experimental constant-expression engine covering more C++ semantics while changing how it represents pointers and integers.
A new System One endpoint serves six decision-model sizes for typed text, JSON, image and video questions, but the code is still nightly-only.
The Q3 Security Signals report treats vulnerabilities, configurations, credentials and sensitive resources as one attack path—shifting priority from isolated scores to reachable outcomes.
A working Muse Spark Open stack shows what operators can use now; broader day-zero model coverage and performance publication remain commitments.
The new recipe preserves the September model’s deployment shape across AMD and NVIDIA hardware, but its published validation still belongs to the older weights.
Three new recipes turn model support into reproducible deployment contracts across AMD text inference, single-GPU speech and NVIDIA image generation.
A Red Hat engineering walkthrough connects policy gates, an SRE agent and human approval into a repeatable failed-build response pattern.
A day-zero image packages text-to-image and multi-reference editing behind OpenAI-compatible endpoints, but the checkpoint is gated and the benchmark is one GB200 run.
The verified recipe separates model-weight fit from KV-cache capacity, with distinct tensor-, expert- and data-parallel choices for B300 and H200 systems.
The new LTS line adds preview semantic routing, fail-closed policy components and a broader MCP and observability surface for agentic integrations.