Optimum Intel v2.1.0

Optimum Intel v2.1.0 is out, alongside three other releases that were co-developed and validated against: OpenVINO 2026.3, OpenVINO GenAI 2026.3, and NNCF 3.3.

New model support

v2.1.0 adds export and inference support for:

• Gemma 3n : Google's efficient multimodal variant of Gemma, pairing a MatFormer-based language model with a MobileNet v5 vision encoder and a USM-based audio encoder.
• Gemma 4 Unified : the encoder-free variant of Gemma 4, which projects raw image and audio inputs directly into the LLM's embedding space through lightweight linear layers instead of using separate encoder towers.
• Qwen3-Omni-MoE : Alibaba's unified multimodal model, built on a Thinker-Talker MoE architecture.
• Qwen3-VL-Embedding : Qwen's vision-language embedding model designed for multimodal information retrieval and cross-modal understanding.
• SmolLM3: a compact, multilingual reasoning-capable LLM, decoder-only transformer using grouped-query attention and NoPE layers for long-context performance.
• FLUX.2 : Black Forest Labs latest series of image generation models.

Check out our documentation for export details specific to each architecture.

Get started

To install the latest release of 🤗 Optimum Intel with the corresponding required dependencies:

pip install --upgrade optimum-intel

We recommend pinning to the versions this release was validated against:

pip install openvino==2026.3 openvino-tokenizers==2026.3 nncf==3.3

Full details on this release are in the v2.1.0 release notes. If you hit an issue or want to contribute support for a new architecture, open an issue or PR.