What changed

OpenVINO 2026.4 was released on 16 September 2026. Intel's release notes list Multi-Token Prediction (MTP) speculative decoding for Gemma 4, Qwen3.5 and Qwen3.6 on CPUs and GPUs. The OpenVINO GenAI component notes additionally list MTP support for Qwen3.5, Qwen3.6 and Qwen3.8. This is a runtime-stack update, not the release of new model weights or a promise of identical acceleration on every machine.

Speculative decoding proposes several next tokens and lets the main model verify them. The developer's claim of lower latency and higher throughput has to be reproduced with the actual model, context length and hardware. The OpenVINO GenAI 2026.4.0.0 package is also available from the project's official repository.

Why a business might care

For an on-premises assistant over internal documents or a customer-request workflow, the release is a reason to measure the cost and speed of an existing model again on available Intel CPUs and GPUs. Retrieval-augmented generation, access control and answer verification remain separate parts of the system: faster generation does not make an incorrect source document correct.

A practical test is to pin the model and driver versions, run the same requests before and after the upgrade, and compare time to first token, time to complete response, p95 under representative load, memory consumption and cost per accepted answer. Start in a test environment, without switching production traffic.

Limitation and image credit

MTP support is limited to the listed models and configurations. Memory requirements, model-weight licensing and Russian-language answer quality depend on the specific model. Do not extrapolate results from one GPU to another without measurement. The photograph shows an Intel Arc A770 as an example of local GPU hardware, not evidence that this card was benchmarked for the 2026.4 release.

Photo: Telaneo, Wikimedia Commons, CC0 1.0; square crop by the editors. [Original image](https://commons.wikimedia.org/wiki/File:Intel_Arc_A770_front.jpg), [CC0 terms](https://creativecommons.org/publicdomain/zero/1.0/).

Primary source: [OpenVINO 2026.4 release notes](https://docs.openvino.ai/2026/about-openvino/release-notes-openvino.html).