openvinotoolkit/openvino 2026.4.1

on GitHub

Summary of major features and improvements

• Improved GroupQueryAttention support for large language models served through the ONNX frontend:
  • Added causal-attention and partial rotary-embedding support to GroupQueryAttention, broadening coverage of modern LLMs.
  • Extended MatMulNBits with bfloat16 support and reordered weight/scale/zero-point layouts for correct decoding of more quantized models.
  • Fixed correctness issues in the sliding-window KV-cache path that could produce wrong results:
    • an out-of-bounds ScatterUpdate index that could cause memory corruption or a driver crash on GPUs,
    • incorrect Gather shape inference for f8e4m3 and other low-precision data types (f8e4m3 Gather is now enabled for the windowed KV-cache).
• Fixed issues:
  • GPU:
    • Fixed wrong or corrupted results when using remote (zero-copy) output tensors with a transpose at the model output under dynamic shapes.
    • Fixed GRU sequence (RNN) models producing incorrect results with dynamic shapes, and corrected blocked-format reorder offsets.
    • Fixed recursive layout/format propagation through Reshape nodes during GPU graph compilation.
  • NPU:
    • ID 233822: Fixed unexpected application crashes on Intel® Core™ Ultra Series 3 processors on Windows that could occur with certain Intel® NPU PV drivers (32.0.100.4300, 32.0.100.4509, and 32.0.100.4512). It is recommended to migrate to OV 2026.4.1 in case the issue is observed with OV releases 2026.2, 2026.3, or 2026.4.
    • Hardened compiled-blob import against malformed or forged blobs: fixed a possible integer overflow in metadata parsing, bounded recursion in the compatibility-string parser, and rejected forged blobs on import.
  • CPU:
    • The OV_CPU_MAX_ISA instruction-set cap now works in release builds (previously debug-only) and is modeled as a feature bitmask.
    • Aligned JIT kernel selection with the OV_CPU_MAX_ISA cap so operators produce correct results when the ISA is capped.
  • Core / Frontends:
    • Hardened operators against malformed inputs: added index validation to ScatterElementsUpdate, divide-by-zero and overflow guards to Col2Im shape inference, and axes validation preventing a potential out-of-bounds write in the ONNX Pad operator.

You can find the OpenVINO™ toolkit 2026.4.1 release here:

• Download archives* with OpenVINO™
• OpenVINO™ for Python: pip install openvino==2026.4.1

Release documentation is available here: https://docs.openvino.ai/2026

Release Notes are available here: https://docs.openvino.ai/2026/about-openvino/release-notes-openvino.html

*Other names and brands may be claimed as the property of others.