Pichai Backs Open Models, Adding Google to a Fight Over Chinese AI

Google’s company-reported benchmark comparison of Gemma 4 models and Gemma 3; the Arena AI text score is dated April 2, 2026. Source: Google Blog.

Google argues from a mixed AI portfolio

Pichai’s position is not simply a general endorsement of openness. Google’s own Gemma 4 announcement says the downloadable model family complements Gemini, its proprietary model line. In other words, the company presents developers with both models they can run or adapt and controlled products it operates itself.

Google says that, since Gemma’s first generation, developers have downloaded the family more than 400 million times and made more than 100,000 variants. Those are company-reported cumulative figures, not an independent measure of usage or model quality. They nonetheless identify a large developer constituency that could be affected if rules limited access to, or work with, particular downloadable models.

The deployment details also matter. Google says Gemma 4 is available under Apache 2.0 and comes in four sizes: effective 2-billion- and 4-billion-parameter edge models, plus 26-billion-parameter mixture-of-experts and 31-billion-parameter dense models. It says the larger unquantized versions fit on one 80GB Nvidia H100 GPU, while quantized versions can run on consumer GPUs; its smaller models are intended for offline edge devices. That range means the policy dispute touches both cloud-scale infrastructure and local deployment, even though the report does not say what technical threshold any U.S. policy would use.

“Open weights” and “open source” are related, not interchangeable labels

Pichai described Gemma as open weight. Google also calls Gemma 4 open and says it uses an Apache 2.0 license. Neither description, by itself, answers every question raised by the Open Source Initiative’s Open Source AI Definition.

The 2024 definition is a standard for assessing whether an AI system can be deemed open source. It says an open-source model must provide enough information about its training data for a skilled person to recreate a substantially equivalent system using the same or similar data. That is a disclosure question beyond the availability of trained weights.

The distinction does not determine whether Chinese models should face security scrutiny. It does, however, prevent the policy debate from collapsing models with different licenses, disclosures, deployment options and alleged conduct into one category.

What evidence would resolve the dispute?

The next decision belongs to the administration: whether to pursue specific alleged IP theft or national-security concerns model by model, as the report indicates officials are considering, or to restrict a broader category of Chinese AI systems.

Before that choice can be evaluated, the public would need the operative standard—what conduct, model capability, distribution channel or security evidence would trigger action. Without it, Pichai’s support for open models and his rivals’ warnings about distillation describe competing preferences, rather than a defined rule with known effects.