A row of national flags. Image: Neil Owen / Wikimedia Commons, CC BY-SA 2.0
Governments around the world are pouring money into “sovereign AI,” homegrown models meant to free them from dependence on American and Chinese tech giants. But a new study from Counterpoint Research finds that most of those national models are built on top of someone else’s: 56% are adapted from existing open-source models, and the most popular foundation by far is Meta’s Llama.
What the study found
Counterpoint’s H1 2026 Sovereign AI LLM Index tracked more than 170 models across over 55 countries outside the U.S. and China, according to the research firm. It found:
- 56% of sovereign models are adapted: fine-tuned versions of existing open-source models.
- 44% are foundational: built from scratch and trained on local data.
Among the adapted models, three foundations dominate:
- Meta’s Llama: 38%, and the top base in every region Counterpoint studied
- Alibaba’s Qwen: 20%, especially popular in Central and Eastern Europe, Asia-Pacific and Africa
- Mistral: 19%, the French lab gaining ground in the Middle East through partnerships like Abu Dhabi’s G42, as well as in Europe and Asia-Pacific
Countries are also building on Google’s Gemma, Nvidia’s Nemotron and DeepSeek, which Counterpoint reads as a sign that nations are spreading their bets rather than relying on a single foundation.
“Control is a continuum”
The findings raise an awkward question: how sovereign is a sovereign model that’s really Llama underneath? Counterpoint research director Marc Einstein doesn’t pretend otherwise.
Sovereign AI is about control and strategic autonomy, but control is a continuum rather than a switch. A slight majority of ‘sovereign’ LLMs are fine-tuned on open-source foreign bases, so true autonomy is still the exception.
Marc Einstein, Counterpoint Research
But he sees the trend moving in the right direction for countries that want independence: “What matters is the direction of travel, and our study shows a clear push towards foundational models built from scratch on native data.”
Why Qwen’s rise matters
The quieter headline is Qwen. A Chinese model is now the foundation for one in five adapted sovereign models, largely thanks to strong support for many languages, which matters to countries building AI in their own. At a time when the U.S. is trying to limit China’s AI progress, Chinese open models are quietly becoming part of other nations’ national AI infrastructure.
It’s also a reminder of how much Meta’s decision to release Llama openly has shaped the global AI map. Building a model from scratch is slow and expensive, and for most countries, the fastest route to a “national” AI has been to start with Meta’s.
Source: Counterpoint Research
