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.
The result is a software-stack comparison, not a clean GPU shootout, but the engineering behind it offers a useful optimization map.
Red Hat’s architecture guide explains why higher VM density depends on coordinated reclaim, tiering, pressure signals and cluster rebalancing—not one overcommit ratio.
New Red Hat AI checkpoints split the performance problem between lower-precision execution and parallel speculative decoding on NVIDIA Hopper GPUs.
A directional five-year model argues that tiering enterprise AI traffic across open-weight and frontier models can cut TCO, but the result depends heavily on workload mix and utilization.
The Q3 report combines Lightwell package observations with an explicitly modeled Java scenario, estimating how shared remediation could avoid duplicated work.
The design separates the delivery of signed fixes for pinned dependencies from the approvals, testing, rollback and audit work needed to deploy them.
The quarterly engineering update emphasizes safer CI, more predictable device-state handling and fixes for large configurations and connection timeouts.
A new engineering comparison says benchmark parity makes traffic shape, hardware support and ecosystem fit the deciding factors.
A 52× HLS speedup for about $26 shows why cache-aware serving may matter as much as model choice in automated design exploration.
A two-shard tech-preview configuration sharply reduced slow applies during event storms, but added memory cost and regressed under one quota-pressure workload.
A new worked example uses Camel’s AI Tool component and built-in MCP server to expose three narrowly defined CRM operations without handing the model arbitrary SQL access.
A new IBM Research explainer connects llm-d’s router, cache tiers and split worker pools to the uneven traffic generated by AI agents.
The field report connects OpenShift failure injection to release gates, virtualization migrations, database recovery and payment SLOs.
Decision 3.0 ships six Apache-2.0 models that answer choice, yes/no and scoring questions with probabilities instead of generated prose.