On 2 September, Multiverse Computing announced Quasar 438B on X as “the top European AI model,” externally validated by Artificial Analysis. There is a Community Note on the post, shown publicly. It reads in full:

its compressed glm 5.2 their business is to make compressed models and have other similar ones but unlike those they are trying to misrepresent this one as a completely new european model made by them

Building on open weights is normal, legal and often sensible. Multiverse’s entire business is model compression — the product is called CompactifAI, and its models catalog opens by describing “our collection of cutting-edge compressed language models.”

Compressing a strong open-weight model and serving it faster and cheaper is a real engineering contribution, and by the numbers this one worked. Artificial Analysis ranks Quasar 13th on intelligence among the 178 models in its comparison class, and measures it at 178.2 output tokens per second. Multiverse also published the gap to the top itself: its own post says the field is “led by Claude Opus 5 at 63,” against Quasar’s 43.

The catalog carries an “Original Architecture” column, filled in for GLM 5.2, GLM 5.3, Nemotron 3 Nano Omni and Whisper Large V3 Turbo Slim. It is blank for Quasar 438B. It is also blank for Hypernova 60B, Carina 60B and Qwen 3.8 27B, the last of which announces its base in its own name. The column is maintained inconsistently across the catalog, so a blank for Quasar carries no weight on its own.

None of this makes Quasar a bad model. It scores 43, it is fast, and at $0.60 per million input tokens it is worth testing if speed and cost are what you are optimising for.

The problem here is the sovereignty claim.

“Europe’s leading AI model” and “Europe’s sovereign AI capability” are not benchmark claims, they are provenance claims, and provenance is what the launch materials leave out.

A European enterprise choosing Quasar to reduce dependence on foreign models is making a decision about where the weights came from, and on that question the launch post does not name a base model and the changelog names GLM 5.2. A benchmark score tells you how the model performs. It does not tell you who trained the weights, and the launch post does not either.

That gap matters more than an ordinary marketing overstatement would, because sovereignty is currently being sold as a category across Europe, on the argument that buyers should prefer local provenance on principle.

If “European model” can mean an MIT-licensed Chinese base with European compression and post-training applied, that is a reasonable product, provided the marketing says so.

Right now a buyer has to find the changelog to know.