pypi / openvino
openvino
What changed in openvino 2026, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-18.
2 security8 deprecations22 new APIs and options2 performance6 behavior and defaults21 fixes and other
See which of these apply to your code, ranked, in your agent:
npx @unstale/cli initopenvino 202661 changes
Security 2
Hardened import against malformed or forged blobs with integer-overflow protection, bounded compatibility-string parsing, and forged-blob rejection.
Added index validation to harden the operator against malformed inputs.
Deprecations 8
Runtime model reshaping through auto batch size is deprecated and will be removed; dynamic shape models are recommended instead.
Runtime model reshaping through auto shape is deprecated and will be removed; dynamic shape models are recommended instead.
Modifying runtime information through read-only ConstOutput objects in the Python API is deprecated; use non-const outputs when updating runtime metadata.
OpenVINO GenAI deprecates finish_chat(); pass a ChatHistory object directly to generate() instead.
The legacy graph iteration path enabled with ONNX_ITERATOR=0 is deprecated and will be removed in future releases.
The dedicated OpenVINO operator for Kubernetes and OpenShift is deprecated in favor of the KServe operator and will no longer be actively developed.
Support for Python 3.10 is scheduled to end in OpenVINO 2026.5 because Python 3.10 has reached end of life.
OpenVINO GenAI deprecates start_chat(); pass a ChatHistory object directly to generate() instead.
New APIs and options 22
Added partial rotary-embedding support for models using GroupQueryAttention.
OpenVINO GenAI adds ASRPipeline support in Node.js, providing a JavaScript pipeline API for streaming automatic speech recognition with performance metrics.
BGE M3 becomes available as an early release with CPU and GPU support.
BGE Reranker-V2-M3 becomes available as an early release with CPU and GPU support.
OpenVINO adds CPU and GPU support for DeepSeek OCR-2.
OpenVINO GenAI previews DFlash acceleration for Qwen models on GPUs.
Fun-ASR-Nano becomes available as an early release with CPU and GPU support.
Granite 4.0 H Tiny becomes available as an early release with CPU and GPU support.
OpenVINO adds CPU and GPU support for Hy-MT2-1.8B.
OpenVINO extends ITT profiling to NPU execution, allowing Intel VTune Profiler to analyze CPU, GPU, and NPU execution through one toolchain.
OpenVINO adds CPU, GPU, and NPU support for Kokoro-82M.
LFM2.5-8B-A1B becomes available as an early release with CPU and GPU support.
MiniCPM5-2B becomes available as an early release with CPU and GPU support.
OpenVINO GenAI adds MTP speculative decoding for Gemma 4, Qwen3.5, and Qwen3.6 on CPUs and GPUs to increase throughput and reduce latency without sacrificing accuracy.
OpenVINO adds CPU and GPU support for Muse Glimmer 30B.
OpenVINO Model Server adds support for Muse Glimmer 30B.
Qwen-Image becomes available as an early release with CPU and GPU support.
OpenVINO adds CPU and GPU support for Qwen3-ASR.
OpenVINO adds CPU and GPU support for Qwen3-VL-4B with EAGLE-3.
OpenVINO Model Server adds support for Qwen3.8 27B.
RF-DETR becomes available as an early release with CPU and GPU support.
OpenVINO GenAI supports Top-K Tree Drafting for EAGLE-3, enabling higher throughput on VLM pipelines than Chain Drafting (Top-1).
Performance 2
OpenVINO Model Server previews idle model management that unloads models when they are not in use to reduce memory usage.
OpenVINO GenAI optimizes Gemma 4 inference with long-context inputs on Xe3 integrated graphics in Intel Core Ultra Series 3 processors.
Behavior and default changes 6
Added causal-attention support for models using GroupQueryAttention.
Preview bounded dynamic-shape support on NPUs extends to vision workloads such as image super-resolution and has been validated with ESPCN.
OpenVINO adds NPU support for FLUX.2-Klein 4B.
OpenVINO adds CPU support for Gemma-3n.
ENABLE_CPU_PINNING is now supported on dual-socket Intel® Xeon® 6+ processor Windows Servers.
Qwen3.8-27B is functionally enabled for use with OpenVINO.
Fixes and other changes 21
Added divide-by-zero and overflow guards to shape inference.
Corrected shape inference for f8e4m3 and other low-precision data types; f8e4m3 Gather is now enabled for the windowed KV-cache.
Fixed incorrect results for sequence models with dynamic shapes and corrected blocked-format reorder offsets on GPU.
Added bfloat16 support and reordered weight, scale, and zero-point layouts to correctly decode more quantized models.
Fixed unexpected crashes on Intel Core Ultra Series 3 Windows systems with certain NPU PV drivers; migration to this release is recommended for affected earlier releases.
Added axes validation to prevent a potential out-of-bounds write.
The instruction-set cap now works in release builds and is modeled as a feature bitmask.
Aligned JIT kernel selection with the ISA cap so operators produce correct results when the cap is set.
Fixed wrong or corrupted results with remote zero-copy output tensors when a model output has a transpose under dynamic shapes.
Fixed recursive layout and format propagation through Reshape nodes during GPU graph compilation.
Fixed an out-of-bounds index in the sliding-window KV-cache path that could cause memory corruption or a GPU driver crash.
OpenVINO adds CPU and GPU support for Gemma 4 12B.
OpenVINO adds CPU and GPU support for Granite 4.0 H Micro.
OpenVINO adds CPU and GPU support for Qwen3.8 27B.
OpenVINO GenAI previews visual-token support to reduce latency and speed up GenAI pipelines on Intel Core Ultra Series 3 processors.
Z-Image-Turbo becomes available as an early release with CPU and GPU support.
The chat template was updated so it can be parsed by the Minja templating engine.
Fixed an accuracy issue when the input prompt is long.
Fixed an accuracy issue when an image is included in the input.
Fixed an accuracy issue for these models when targeting an NPU device.
Muse-Glimmer-30B is functionally enabled for use with OpenVINO.