CONFIRMED

Europe just got a measured workhorse, not a flagship. On September 2, 2026 San Sebastián's Multiverse Computing published Quasar 438B, the lab's first large model, aimed at enterprise agents and coding. It runs English and Spanish, lists 438 billion parameters, keeps reasoning on, and ships through the CompactifAI API. The 'leading European model' line is marketing on a thin field: Quasar scores 43 on the Artificial Analysis Intelligence Index, ahead of Mistral Medium 3.5 (30) and Nemotron 3 Ultra (38). Claude Opus 5 sits at 63 on the same index. Anyone who needs Portuguese, vision or open weights is still outside this product.

What the independent card actually shows

Artificial Analysis confirms the 43, ranks Quasar 13th of 178 models in class, and describes a proprietary, text-only system with a 1 million token context. Speed: 176.2 tokens per second. First-party API pricing: $0.60 input and $1.80 output per million tokens. The lab says the model returns 500 tokens, thinking included, in 15.3 seconds, and scores 75.0 on AA-LCR, level with Grok 4.6 at high. On Terminal-Bench v2.1 their figure is 69.3, 18.7 points above Mistral Medium 3.5 and well behind Opus 5 at 89.1. That changes the practical reading: Quasar fights the European and cheap-American mid table, not Fable, Sol or Google's new Flash. AA still lists the model as released on August 10. The September 2 product post is the commercial announcement, not necessarily the first day the endpoint existed.

Two-language sovereignty and a closed API

The European pitch Multiverse sells is choice plus latency. Cofounder and CEO Enrique Lizaso says the company does not have to pick between reasoning and speed. The actual cut is narrower. No open weights. No public architecture card, training data or active parameter count. The community has already asked whether this is a third-party fine-tune; official docs list Quasar as an original model. Without training evidence, that remains the company's claim. The model does not take images. It does not speak Portuguese. The European Commission is pushing another line, open models across the Union's 24 languages. For a Brazilian SME, Spanish helps if the client is Hispanic. It does not replace Qwen3.8-Max, Gemini 3.8 Flash or Grok 4.6 in local production.

What changes in the agent bill

$0.60 and $1.80 is cheaper than Grok 4.6 at $2 and $6, and more expensive than the AA floor (median $0.25 and $0.90). AA itself calls Quasar verbose: 350 million output tokens on the index against a 67 million median. A long agent that plans, calls tools and corrects itself can eat the price edge. Run a short test on terminal coding and on a long English or Spanish dossier, with a token cap. Do not migrate a stack, do not buy European sovereignty as a slide, and do not treat 43 as frontier. Qwen3.8 27B, with far fewer parameters, ties 43 on low and climbs on xhigh. The continental ranking is true inside AA's narrow cut. It is not the ranking that picks a marketing, support or product tool in Brazil.

MaxAssistant reading

The signal is real and labeled. Confirmed: Multiverse put a 438B on the API, Artificial Analysis scored 43, the price and the 1M window exist. Hypothesis: the Spanish compression house wants to become a European agent vendor while Mistral stalls on Medium. Feeling: the announcement lands on a day Google, Qwen and Anthropic already moved the desk, so Quasar enters as a niche option, not a market event. Test it now. Buy the leading-in-Europe story, no.

Sources

Multiverse Computing, Introducing Quasar 438B: Europe's Leading AI Model: https://multiversecomputing.com/resources/introducing-quasar-438b-europe-s-leading-ai-model | Artificial Analysis, Quasar 438B: https://artificialanalysis.ai/models/quasar-438b | The AI Insider, Multiverse Computing Launches Quasar 438B: https://theaiinsider.tech/2026/09/02/multiverse-computing-launches-quasar-438b/ | Frandroid, Quasar 438B, the new European leader that only speaks two languages: https://www.frandroid.com/culture-tech/3230553_quasar-438b-le-nouveau-leader-de-lia-europeen-depasse-mistral-sur-les-benchmarks-mais-ne-parle-que-deux-langues