A Spanish startup just put Europe's highest-scoring model on Artificial Analysis' Intelligence Index, and it did it by compressing an existing model instead of training a bigger one from scratch.
Multiverse Computing launched Quasar 438B on September 2, 2026. It's the company's first large language model. By one closely watched measure, it's also the strongest model to come out of Europe. The San Sebastián based company says Quasar scored 43 on the Artificial Analysis Intelligence Index v4.1.1: a benchmark blending nine tests across agent tasks, coding, scientific reasoning and long-context knowledge. Forty-three points is not nothing. That's 13 points ahead of Mistral Medium 3.5, the French model many people treat as Europe's default answer to OpenAI and Google. It also beats Nvidia's Nemotron 3 Ultra by 5 points - even though the Nvidia model carries 112 billion more parameters. That's the headline.
You don't usually see a smaller model beat a bigger one from a chipmaker with Nvidia's compute budget. That's the actual story here, not just another leaderboard entry.
Quasar handles English and Spanish, runs on a one-million-token context window, and generates 500 tokens, including reasoning time, in 15.3 seconds, according to figures published by Multiverse and benchmarked through Artificial Analysis. Multiverse says only three models in the comparison respond faster. Just one of those, Gemini 3.7 Flash, also outscores Quasar on intelligence. It's live now through Multiverse's CompactifAI API. The company is pitching it at software engineering, operational automation and document-heavy research work, the kind of enterprise task where a slow model burns money on every API call.
Compression, Not Scale #
The trick isn't scale. It's compression. Multiverse built CompactifAI, a system that uses tensor network methods borrowed from quantum physics research. It shrinks existing large models by 80 to 95 percent while keeping most of their accuracy intact, according to the company. Multiverse hasn't disclosed which base model it compressed to build Quasar, or how much smaller it made it. That makes the comparison harder to check from the outside. But the bet is clear: instead of racing Nvidia and the hyperscalers to buy more GPUs, squeeze more intelligence out of the compute you already have.
Microsoft's MAI-Image-2.5-Pro Just Topped an AI Image Editing Leaderboard Microsoft's MAI-Image-2.5-Pro debuted at number one on Artificial Analysis's Image Editing Leaderboard, beating GPT Image 2 and Reve 2.1 less than a month after its July 23 preview launch on Microsoft Foundry. The model prices at roughly $108.50 per thousand images, more than double its own mid-tier sibling. - microsoft mai image 2.5 pro leaderboard ranking - ai image editing model performance comparison 2026
That's a real strategic difference, not just marketing. Iberdrola, Bosch and the Bank of Canada are already among Multiverse's more than 100 customers. That's real revenue. The company has offices in the US and Canada, plus sites across Europe, on top of its Spanish headquarters. For the European AI sector, that matters. It has spent the last few years watching the big AI labs, OpenAI, Anthropic, Google, pour tens of billions into training runs. A compression-first approach is one of the few paths that doesn't require matching Silicon Valley dollar for dollar.
The Agent Claim Needs Care #
Not everyone is convinced the headline number tells the full story. The New Stack pushed back on Multiverse's framing that Quasar is fast enough for AI agents, pointing to Artificial Analysis' output speed of about 183 tokens per second. Raw benchmark speed doesn't automatically translate into a fast agent once you account for repeated reasoning steps, tool calls and result checking in a live workflow. That's a fair caveat. A model that looks quick on a single 500-token test can still slow down once it's chained into a multi-step agent loop. Multiverse hasn't published agent-specific latency numbers to settle the question either way.
Still, the Intelligence Index result stands on its own, and it's an external one. Artificial Analysis is an independent benchmarking firm, not Multiverse's own marketing team. Its methodology combines GDPval-AA v2, τ³-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience and AA-LCR into a single weighted score. That's a demanding mix. Getting to 43 on that combination, ahead of an Nvidia model with a third more parameters, is a real technical result. Whatever caveats apply to the speed claims layered on top of it, that number still counts.
Europe has talked about sovereign AI for years without much to show for it beyond policy papers and export-control anxiety. Quasar 438B is a small, checkable data point instead. It shows a European lab can compete near the frontier of model intelligence without duplicating the compute budgets of OpenAI or Nvidia. That part is real. Whether Multiverse can keep that edge is the open question. Bigger labs are already working on their next generation of models, and nobody, including Multiverse, has answered that yet.
Also read: Songcove Turns Your Actual Love Story Into the Gift for Wedding Anniversary You Cannot Buy in a Store • Broadcom Says AI Chip Revenue Will Hit $115 Billion by 2027 • Anthropic Open-Sources Claude Commerce Agents Blueprint Ahead of Black Friday