Optimum-Intel v2.2.0 & OpenVINO GenAI 2026.4.0: What's New

Optimum Intel 2.2 is here, alongside OpenVINO 2026.4, OpenVINO GenAI 2026.4 and NNCF 3.4. These releases expand support for language, vision, speech, and image and video generation, while adding new decoding options and more detailed performance metrics.

These libraries cover different steps in deploying a model:

• Optimum Intel exports your model to the OpenVINO format
• NNCF applies quantization on your converted model, to help lower memory and compute requirements
• OpenVINO executes the optimized model efficiently on Intel CPUs, GPUs, and NPUs
• OpenVINO GenAI provides APIs for generative AI tasks, including text generation, visual understanding, image generation, and speech recognition

Together, they provide a complete workflow, taking a model from the HF Hub to an application running inference locally.

What's new in Optimum Intel v2.2

Optimum Intel 2.2 expands model support across several tasks, letting you export new models to OpenVINO's Intermediate Representation (IR) for efficient inference.

Language and vision-language models

• Mistral 3 : supports text and image understanding, including answering questions about images.
• DeepSeek-OCR-2: extracts text and document structure from images, including conversion to Markdown.
• Muse Glimmer: combines language and visual understanding for multimodal assistant and agent applications.
• Ouro: reuses transformer layers across multiple passes, increasing computational depth without adding a separate set of parameters.
• Gemma 4 and Gemma 4 Unified now accept video as input, in addition to text, images, and audio.

Image and video generation

• Qwen-Image : generates images from text prompts, with particular emphasis on rendering text within images.
• Z-Image: generates images using a diffusion transformer that processes text and image representations in a shared sequence.
• LTX-2 : generates synchronized video and audio in a single pass.

Speech

• Fun-ASR: converts speech into text for transcription applications.

Faster generation: speculative decoding

Speculative decoding speeds up text generation by having a small, cheap model propose several tokens at once, which the full model then verifies in a single pass instead of generating them one at a time.

• DFlash draft models are now supported for Qwen3, Qwen3.5, Qwen3.6, Qwen3-Coder, and Gemma 4, including the extra hidden-state export needed to drive them from vision-language targets.
• Multi-Token Prediction (MTP) heads, an alternative to a separate draft model, where the main model predicts several future tokens itself, are now exported for Qwen3.5, Qwen3.6 (dense and MoE), and Qwen3.8-27B, so OpenVINO GenAI can use them for self-speculative generation.

Check out our documentation for the full list of supported architectures.

What's new in OpenVINO GenAI 2026.4

OpenVINO GenAI 2026.4 brings these capabilities into task-specific pipelines.

• New pipelines: image generation for Qwen-Image and Z-Image Turbo, and speech recognition for Fun-ASR-Nano-2512 in ASRPipeline.
• VLMPipeline extended to DeepSeek-OCR-2, Meta Muse Glimmer, Qwen3.8, and video input for the Gemma 4 family.
• Paged Attention for the Mamba 2 layer, enabling granite4.0-h-micro and granite4.0-h-tiny/small on CPU and GPU. Paged Attention is the memory-management scheme that lets a server batch many requests together efficiently; extending it to Mamba 2's recurrent state means these hybrid models can now benefit from the same serving efficiency as regular transformers.
• Faster generation: DFlash speculative decoding extended to Qwen3.6 and to VLMPipeline; Multi-Token Prediction (MTP) speculative decoding added for Qwen3.5, Qwen3.6, and Qwen3.8; Eagle3 top-k decoding enabled for vision-language models.
• Better observability: VLM performance metrics now report vision-encoding, audio-encoding, and text-embedding durations, and continuous batching exposes per-request VLM metrics through GenerationHandle.
• Qwen3-Omni improvements: vision self-attention and image preprocessing can now be offloaded to GPU, and the Talker speech API gains configuration accessors, a property-based generate() overload, and a ModelsMap constructor for blob deployment and per-submodel device placement.
• Node.js: the GenAI Node.js API now includes ASRPipeline, so speech-recognition models can be used from JavaScript/TypeScript projects too.

Get started by installing Optimum Intel and the companion library versions recommended for this release:

python -m pip install --upgrade \
  optimum-intel==2.2.0 \
  openvino==2026.4 \
  openvino-genai==2026.4 \
  openvino-tokenizers==2026.4 \
  nncf==3.4

Quick start

In Optimum-Intel v2.2, we introduced Mamba 2 support in Optimum Intel. The Mamba 2 Selective SSM representations in the OpenVINO IR have been finalized for the Granite-4.0-H model and will serve as a base for supporting future Mamba 2 models.

Granite-4.0-H-Micro can be exported and quantized to int8 with the following command :

optimum-cli export openvino -m ibm-granite/granite-4.0-h-micro --weight-format int8 granite-4.0-h-micro

Inference can be performed with OpenVINO GenAI :

import openvino_genai as ovg

model_path = "./granite-4.0-h-micro-int4"
pipe = ovg.LLMPipeline(model_path, "CPU")
chat_history = ovg.ChatHistory([{"role": "user", "content": "What is the capital of France?"}])
generation_config = ovg.GenerationConfig(max_new_tokens=10, do_sample=False)
output = pipe.generate(chat_history, generation_config=generation_config)