#OpenVINO 2026.4 now features tree drafting (top-k) support for EAGLE-3 to help developers balance richer multimodal experiences without limiting performance. Rather than delivering outputs sequentially, tree drafting enables higher throughput for VLM pipelines compared with chain drafting (top-1). Try OpenVINO 2026.4 for your VLM today at http://ms.spr.ly/6047ac4gF
Intel Devs
IT Services and IT Consulting
From the data center to the edge, Intel Devs is your hub for transforming hardware into possibility.
About us
Connecting the worldwide community of developers on all things software and hardware.
- Website
-
http://software.intel.com
External link for Intel Devs
- Industry
- IT Services and IT Consulting
- Company size
- 10,001+ employees
Updates
-
Intel Devs reposted this
Real-Time Vision Applications with RF-DETR and OpenVINO™ 2026.4, now running on Intel hardware. No convolutional backbone, no non-maximum suppression cleanup step: this DETR-style detector from Roboflow predicts its boxes directly, and as of OpenVINO 2026.4 it runs on Intel CPUs and GPUs. Anna Gubenkova , Dmitriy Pastushenkov and I put together a walkthrough on how to run it. If you're building anything where detection latency needs to stay predictable, robotics, retail cameras, quality-control lines, this is worth a look. 🔗 Full walkthrough here: https://lnkd.in/gy7KaRQZ #OpenVINO #ComputerVision #RFDETR #EdgeAI #IntelDeveloper
-
Running LLMs on Intel Xeon is easier than ever with Intel Inference Microservices. Built to help developers serve LLMs through a single Docker container, Intel Inference Microservices is the optimized building block for getting LLMs up and running quickly and efficiently. OpenAI-compatible API endpoints let developers choose a model, skip tuning configurations, and jump straight to optimized deployment by default. Download it today: http://ms.spr.ly/6049ac4iV
-
-
RF-DETR now has early release support in OpenVINO and the model can be exported to OpenVINO model format directly! Learn more here: https://lnkd.in/gQkk93C7
Lead maintainer of RF-DETR · Maintainer of supervision @ Roboflow · Co-creator of PyTorch Lightning & TorchMetrics · PhD in Medical Imaging · Kaggle Master
#RFDETR v1.11.0 is out. We made it easier to ship and quicker to train. The carousel walks through the numbers; here is the short version. 📦 Deploy where you already run - #OpenVINO IR with float32/float16 precision - #LiteRT straight from torch.export, no #ONNX/#TensorFlow hop (experimental) - Apple #CoreAI (.aimodel) for iOS/macOS 27+ (experimental) - Dynamic-batch #TensorRT: one engine for batch 1..max instead of one per size. You give up some speed away from the tuned batch and ship a single file. - Opt-in Inductor backend for long-running inference at a fixed batch size and resolution ⚡ Train faster, fit more - #PyTorch compile: 1.50x eager throughput on an NVIDIA #RTX6000 (batch size 128) and 1.41x on an #L4 (batch size 32) - #FP8 through Transformer Engine: 8-13% less peak memory than compiled BF16. Same speed on the RTX6000, 8-11% slower on the L4. - Opt-in CUDA-graph replay for small batches: -32% median step on an L4 at batch size 4 - #WebDataset tar-shard streaming and simplejpeg decoding to keep the GPU fed 📊 Pick your COCO evaluator with one flag - #vernier (new default), #hotcoco and #ufcoco, all Rust, plus #faster_coco_eval, the evaluator used before v1.11.0 Thank you to the 13 contributors who made this release, credited in the comments 👇 If you deploy #ObjectDetection or #ComputerVision models to a target the exporters don't cover yet, the #OpenSource repo takes issues and PRs.
-
What changes when your AI model has to work in the physical world? Now you’re connecting cameras, robots, data collection, training, benchmarking, and inference into one working system. Guy Tamir explores that full workflow with Physical AI Studio, from connecting the environment and collecting data to training, benchmarking, and running models. For developers getting into physical AI, it’s a practical look at how the pieces fit together and what it takes to move from a model to a working robotics system. Check out Physical AI Studio: http://ms.spr.ly/6048acuyn
-
"Local or cloud?" For agentic workloads, more teams are answering "both, depending on the request." This on-demand session breaks down how Intel AI Builder's hybrid router makes that call automatically, and how #OpenVINO Model Server(OVMS) serves the local half efficiently across Intel hardware, from AI PCs to edge devices. This session covers: 🔹Rapid creation of custom AI assistants and agents tailored to specific industry needs with Intel AI Builder. 🔹A benchmark showing near-cloud quality with most tokens processed locally. 🔹Live deployment and agent demos, including OVMS-served models wired into Cline coding agent and an MCP-connected tool call. Worth the watch if you're weighing where your next agent should actually execute. Watch on demand: http://ms.spr.ly/6044a9aBo
-
Intel Devs reposted this
Physical AI Studio v0.2.0 is out! This release brings support for new policies: RLDX-1, MolmoAct2 and XR0, plus LoRA/DoRA fine-tuning for Pi0 and Pi0.5, a guided training setup wizard, one-click image augmentation, and remote training on AWS straight from the Studio UI. What I like most: it's genuinely easy to install and use. We just tried it on an Intel NUC and had it up and running in minutes. This is exactly why tools like this matter for Ignite Next and our hackathons. When the barrier to entry is low, talented people can spend their time solving real problems instead of fighting with setup. So here's an open invitation: if your company has a real-world robotics or Physical AI challenge, bring it to us. We'd love to put it in front of hackathon talent and see it tackled with Physical AI Studio and the Physical AI library. 🔗 Release notes: https://lnkd.in/dB8qxuMX 🔗 Physical AI Studio: https://lnkd.in/dVeU3Ppz 🔗 Physical AI library: https://lnkd.in/dcHj8QGW Huge thanks to the team and contributors for this release! 👏 Radwan Ibrahim Samet Akcay Alfie Roddan Daniil Lyakhov Mark Redeman Alexander Barabanov Vladislav Sovrasov Albert van Houten Daan Krol CC: Olinca Beltran Sales Jayabalaji Sathiyamoorthi Alois Eder Alexey Moskalev Ilya Efimov Yannis Katramados Guy Tamir #PhysicalAI #Robotics #OpenSource #Intel #OpenVINO #Hackathon #IgniteNext
-
Public clouds aren’t the only way to bring modern AI to your enterprise. Don’t trade data sovereignty and compliance for performance—Deloitte and Red Hat engineered a private cloud architecture to keep your sensitive workloads where they belong. Built on #IntelXeon 6 processors, Red Hat OpenShift, and Red Hat Enterprise Linux CoreOS, this solution anchors governance through confidential computing. Secure your IP without slowing deployment. Learn more at http://ms.spr.ly/6045a9zqf
-
-
When you're creating a federal AI workflow, size, weight, power, connectivity, and reliability are critical constraints. The Federal and Aerospace AI Suite inside #OpenEdgePlatform 2026.2 is built to help modern defense and aerospace systems deploy sophisticated AI in those restricted environments. Complete with SWaP-optimized mission AI, the Federal and Aerospace AI Suite uses pre-validated blueprints and development patterns optimized for tactical edge environments so that organizations can execute complex tasks in a single-SoC workload, completely disconnected from the cloud. Secure your AI objectives. Download the Federal and Aerospace AI Suite today: http://ms.spr.ly/6045agqER
-
-
Estimating the needs for your AI efforts? The new #IntelXeon Processor AI Advisor can help you identify the optimal configuration for your workloads. Instead of navigating logistics like model size, latency requirements, concurrency, and costs, let our tool consolidate those major details into a single experience so you can pinpoint the platform, stack, and configuration for your SLA. Learn more and try the Intel Xeon Processor AI Advisor for yourself: http://ms.spr.ly/6042agqG6
-