Optimum Intel 2.2 and OpenVINO GenAI 2026.4 add DFlash and multi-token-prediction speculative decoding for recent Qwen and Gemma models, including vision-language paths. The suite also brings paged attention to Mamba 2 hybrids, adds image-generation and speech-recognition pipelines, and exposes per-request vision, audio and embedding timing. It expands local CPU, GPU and NPU support across Mistral 3, DeepSeek-OCR-2, Qwen3.8, Gemma 4 and other multimodal models.
OpenVINO 2026.4 brings speculative multimodal serving to Intel hardware
2026-09-17Verified announcementTRIAL
Optimum Intel 2.2 and OpenVINO GenAI 2026.4 add DFlash and multi-token-prediction speculative decoding for recent Qwen and Gemma models, including vision-language paths. The suite also brings paged attention to Mamba 2 hybrids, adds image-generation and speech-recognition pipelines, and exposes per-request vision, audio and embedding timing. It expands local CPU, GPU and NPU support across Mistral 3, DeepSeek-OCR-2, Qwen3.8, Gemma 4 and other multimodal models.
Release noteshuggingface/optimum-intel · releases / tag / v2.2.0Release notesopenvinotoolkit/openvino.genai · releases / tag / 2026.4.0.0
Availability: Apache-2.0 open-source packages; production release
Imported from the existing research archive. Migration preserves the recorded judgment and is not a fresh verification.
What this means for you
For teams deploying local or edge inference on Intel hardware, this closes several gaps that previously pushed multimodal and hybrid models toward CUDA-centric stacks. Trial it on one production-shaped model and measure end-to-end latency, memory and export fidelity; the release publishes capabilities, not cross-hardware benchmark results.
Evidence & related material
Technical details & scores
| Release Suite | Optimum Intel 2.2.0, OpenVINO 2026.4, OpenVINO GenAI 2026.4 and NNCF 3.4 |
|---|---|
| Speculative Decoding | DFlash for Qwen3/3.5/3.6, Qwen3-Coder and Gemma 4; MTP export/runtime paths for Qwen3.5, Qwen3.6 and Qwen3.8-27B; Eagle3 top-k for VLMs |
| Hybrid Serving | Mamba 2 paged attention for Granite 4.0 H models on CPU and GPU |
| Pipelines | Qwen-Image and Z-Image generation; Fun-ASR speech recognition; expanded VLM video and OCR support |
| Observability | per-request vision-encoding, audio-encoding and text-embedding durations |
| Availability | Apache-2.0 open-source packages; production release |
Technical impact 4
Engineering value 4
Adoption signal 4
Maturity 4
Verification cost 3
Risk 2