Europe's top-scoring AI model is a compressed GLM-5.2
Spain's Multiverse Computing has spent two weeks selling Quasar as proof that Europe can field a frontier model of its own. The company's own documentation says the foundation underneath it was trained in China.
Multiverse Computing released Quasar 1.1 438B on September 15 — a rebuild of the model it launched on September 2 as the highest-scoring European entry on the Artificial Analysis Intelligence Index — and its own write-up states that Quasar is built from GLM-5.2, the open-weights model from China's Z.ai, run through the company's CompactifAI compression pipeline. The rebuild prunes every mixture-of-experts layer from 256 experts down to 148, then re-heals the pruned network on reasoning traces, tool-call sequences and general knowledge. Multiverse reports reasoning gains (HLE up 6.2%, GPQA up 4.3%) and instruction-following gains (IFBench up 4.6%), with average output falling from 3,322.9 tokens to 2,074.3 — 37.6% fewer — across five coding, reasoning and question-answering benchmarks.
Part of the healing set was generated by a hybrid model with a sixth of its layers replaced by a quantum neural network, with circuits run on a 156-qubit IBM Heron processor at IBM Quantum System Two in San Sebastián. The company also stripped what it calls inherited refusal behaviour: refusal on politically sensitive prompts drops from 71.18% on GLM-5.2 and 63.75% on Quasar 1.0 to 41.00%, while refusal on harmful prompts holds at 93.00% on JailbreakBench. Quasar is API-only, sold through Multiverse's CompactifAI service, and priced above the median of its peer group.
The lineage was never hidden in the paper. It was missing from the promotion. Chinese technology press picked up the gap after a post on X drew close to 400,000 views: developers noted the launch thread had room for a European flag and a quantum processor but not for the base model, and described Quasar as "a compressed GLM-5.2 dressed up as a brand-new European model." Artificial Analysis labels the entry "Quasar 438B (max, based on GLM-5.2)" and dates it to August, ahead of Multiverse's September 2 announcement. Compressing and serving someone else's weights is real engineering, and a 43 on Intelligence Index 4.1.1 is a real number. Calling the result European leadership is a procurement argument, not a technical one — and it is the same argument Brussels has been funding since Draghi told Europe to build AI data centres while holding about 5% of global compute.
Beijing's Naive AI has raised about $400 million across three rounds in seven months and now carries a valuation of just over $1.4 billion — before shipping a single model. The Information first reported the rounds; Chinese outlets add the breakdown — $100 million, then $180 million, then $120 million — with Tencent, IDG Capital, MPCi and Sequoia China among the investors. The company was founded in February by Dai Jifeng, a Tsinghua associate professor known for deformable convolution networks and the InternVL multimodal family, and has stayed almost invisible: its site shows a slogan and nothing else, and its X account opened this month with no posts.
The plan is to release its first large model as early as this month, as open weights that anyone can download and customise — and it will not be trained from scratch. Sources told The Information that Naive AI is building on an existing domestic open-weight model, a route Mira Murati's Thinking Machines took when it said its Inkling model largely followed DeepSeek's V3 architecture. For a company under 100 people, that is the cheap path to a credible first release; it also means the differentiator will be post-training and agent work, not pretraining scale.
What to watch: whether Multiverse's next release names its base model in the launch post, and whether Naive AI's first checkpoint ships this month as open weights.
If a model is a compressed copy of someone else's weights, does the label "European" or "Chinese" tell you anything a buyer actually needs? Tell us in the comments.
