Intel has released OpenVINO 2026.3, updating its open-source AI inference toolkit with broader model support, new generative AI pipelines, memory-efficiency improvements and expanded hardware compatibility.
The release adds support for SmolLM3-3B, LFM2-1.2B and LFM2.5-1.2B across Intel CPUs, GPUs and NPUs. Additional models supported on CPUs and GPUs include Harrier OSS-v1-0.6B, Qwen3-8B with EAGLE-3, MiniCPM5-1B and FLUX.2-klein-4B. YOLO26 support has also been extended to Intel GPUs and NPUs, while several other models are available as early releases. Compatibility with Hugging Face Transformers 5.5 is included.
OpenVINO GenAI now extends its EAGLE-3 speculative-decoding pipeline to large language models and vision-language models. The update adds Top-K sampling and builds on continuous batching to improve token-generation performance across CPUs, GPUs and NPUs. Three new GenAI pipelines support multimodal Omni workloads, automatic speech recognition and multimodal embedding generation.
For memory-constrained deployments, OpenVINO 2026.3 introduces disk offloading for mixture-of-experts models. Intel says this allows 30-billion-parameter MoE models such as Qwen3-30B-A3B to operate on systems with 16 GB of memory. Linux support for lazy weights is also new, reducing peak memory use by avoiding unnecessary duplication of model data. The Neural Network Compression Framework adds FP8 quantization for ONNX models.
The release introduces support for Intel Xeon 6+ processors. OpenVINO Model Server gains unified REST endpoints and a standard v1/chat/completions endpoint, alongside updates intended to improve LLM serving stability, performance and tool-parser support.
The changes target inference deployments across edge, local and cloud environments, with particular emphasis on running newer generative and multimodal models efficiently on heterogeneous Intel hardware. Developers can use the toolkit to optimize and deploy models across supported processor types while maintaining a common development framework.
Further information is available in the OpenVINO 2026.3 release notes. Developers can also review the system requirements and download OpenVINO from Intel.
