Nvidia hands European AI builders more open weights, but keeps the GPU lock-in

Nvidia released a coordinated open-weights model family on October 20, 2026, headlined by Nemotron 3.5 Embed at 8.38B and 1.55B parameters, an 8.38B Rerank model Nvidia says tops the ViDoRe v3.1 leaderboard, and a 631M-parameter Parse model for document extraction. Nvidia's developer blog frames the release as part of Europe's "sovereign intelligence" push, saying the open models give organizations a foundation they can customize with local data and deploy on infrastructure they control.

Open weights let teams inspect, fine-tune, and run the models on their own infrastructure, which is the part European builders can act on directly. The license is OpenMDW-1.1, which adds Principle of Least Privilege dataset access on top of the standard Apache 2.0 terms that cover the base Ministral-3 components. "Open" is conditional, not unconditional.

The other half stays closed. The NVFP4 quantized Embed 1B variant is built for NVIDIA Model Optimizer v0.47.0, and its model card states that the checkpoint "requires hardware with NVFP4 support, such as NVIDIA Blackwell GPUs," listing Nvidia Blackwell as the supported hardware microarchitecture and reference configuration. The efficient four-bit path is a Nvidia-silicon path.

Independent confirmation of the ViDoRe v3.1 SOTA claim is still pending; the figure is vendor-self-reported on the model card.

The result is a partial answer: European builders gain real agency on the model layer, and the compute question is unchanged.