Red Hat AI

Agentic AI is cool, but how can we make it actually do things for us? Welcome to the world of Computer Use, where your model can take action for you on your desktop, and perform actions like typing, clicking, and more.

In this demo, we run an open source, natively multimodal model (Qwen3.6, 35B sparse MoE) locally on Mac using MLX, then connect it to Hermes, an AI agent with built-in guardrails for computer use. Watch it move a chess piece, apply a Photo Booth filter, and read a graph straight out of a research paper, all by seeing the screen and clicking, typing, and navigating like a human would. Everything runs on your own hardware through an OpenAI-compatible endpoint, so nothing you see or do ever leaves your machine.

#ComputerUse #AIAgents #RedHatAI #LocalLLM #PrivateAI #OpenSourceAI

Learn how to build an open cloud native architecture for AI agents. This blueprint explains how to improve workload isolation and implement inference routing.

Learn how to set up local agentic AI computer use. Run quantized models like Qwen 3.6 with Hermes to automate desktop tasks on your own terms today.

Upgrade your PyTorch training and inference with TorchComms, the new back-end layer for modern large-scale ML systems.

Learn how OpenShift and OpenShell combine for dual protection of AI coding agents, reducing attack surface and securing your infrastructure.

Deploy a self-hosted AI coding assistant with vLLM and Red Hat OpenShift AI for privacy and operational independence.

Discover how to configure EvalHub evaluation collections on Red Hat AI. Run Lighteval, Garak, and GuideLLM in parallel for a unified LLM pass/fail verdict.

Learn how smarter data generation strategies can reduce the cost and time needed to train high-quality speculator models for speculative decoding. This post shares findings on cross-distillation, training efficiency, and production inference gains that deliver faster LLM serving with no loss in output quality.

Learn how to isolate AI agents using the supervisor pattern and OpenShell sandboxes. Protect credentials and limit blast radius during incident response.

Explore a demo of serving a multimodal model (Qwen3-Omni) with vLLM-Omni on a single hardware accelerator.

Learn how to optimize deployment of vLLM for various traffic shapes, including high-concurrency chat, long-context RAG, high-throughput batch, and distributed AI-grid.

Learn about the three optimization levers for distributed AI inference: prefill/decode disaggregation, KV cache strategy, and speculative decoding.

Learn how Red Hat's SastAI initiative, in collaboration with NVIDIA, automates false positive identification in static application security testing (SAST) using generative AI. By employing an agentic, multi-stage research workflow, SastAI reduces noise and improves triage efficiency. Discover the pattern harvesting methodology that greatly enhances the SastAI solution, now offering a tighter solution with better knowledge and reasoning.

Learn how to connect the EvalHub runtime to internal or external model servers using service account tokens, API keys, or custom certificates.

Learn about the five-dimensional design space in modern LLM serving, including tensor, pipeline, expert, data, and context parallelism.

Look inside Red Hat AI Inference on Amazon EKS to understand its core architectural components and Kubernetes resources.

Discover how to use EvalHub and OCI persistence to make your AI evaluation results immutable, content-addressable, and fully auditable.

Explore the mechanics of gradient synchronization in PyTorch distributed training, focusing on MPI primitives like All-Reduce and core techniques like pipeline parallelism, tensor parallelism, and sharded data parallelism.

Learn how speculative decoding can improve the performance of large language models (LLMs) in production by using a small, fast model to generate tokens speculatively and a large model to verify them.

Learn how to use the EvalHub CLI to automate AI evaluations in your CI/CD pipelines. Install the SDK, configure profiles, and set up a production gate.

Learn how llm-d routes each inference request to the GPU that already has the relevant data cached, cutting down on time-to-first-token, and doubling throughput without changing hardware. Discover how Red Hat's stack packages this neatly into a single Kubernetes resource.

Learn how to onboard a custom evaluation framework into EvalHub using one class, one method, and a container image. This guide covers the contract, data structures, and a complete minimal adapter.

Headed to WeAreDevelopers World Congress Europe 2026? Visit the Red Hat Developer booth on-site to speak to our expert technologists.

Learn how to read an existing system collection, understand its threshold logic, and build your own collection that encodes your actual measurement strategy with thresholds that mean something.

Speculators v0.5.0 introduces DFlash support, enabling single-pass draft token generation with block diffusion for more efficient speculative decoding workflows. The release also adds unified online and offline training through vLLM’s native hidden states extraction system, improving training flexibility, version stability, and production readiness.