Spanish AI company Multiverse Computing has officially launched its flagship reasoning model Quasar 438B, scoring 43 points on the Artificial Analysis Intelligence Index v4.1.1. This makes it the highest-scoring AI model in Europe, surpassing Mistral Medium 3.5 (30) and NVIDIA Nemotron 3 Ultra (38). The release marks a substantial step forward for European sovereign AI development.
Performance Positioning: Strongest in Europe, But Still Not at the Global Top Table
Quasar 438B has 438 billion parameters, positioning it for enterprise agents, software development, and complex multi-step tasks. Its 1 million-token context window supports processing of lengthy documents such as contracts, technical manuals, and research reports. Its intelligence score outpaces all European rivals across nine comprehensive benchmarks.
However, the global perspective is more nuanced. A score of 43 is on par with the low-reasoning variant of Alibaba's Qwen3.8 27B, a model with only 27 billion parameters. In contrast, benchmarks like Anthropic's Claude Fable 5.1 now exceed 66 points. Thus, "Europe's best" is less a triumph and more a reflection of Europe's current AI status—possessing talent and compute, but still playing catch-up in the large-model race.
Speed and Long Context: A Differentiated Advantage for Enterprise Agents
Quasar 438B excels in response speed. End-to-end generation of 500 output tokens takes just 15.3 seconds (including reasoning time), with an output speed of 176.2 tokens per second and a first-token latency of only 1.08 seconds. In enterprise agent scenarios requiring multiple model calls, this speed can significantly reduce cumulative delays.
Long-context reasoning is another strong point. Quasar 438B scores 75.0 on the AA-LCR benchmark, matching Grok 4.6 and trailing Claude Opus 5 by only about 0.7 points. On the Terminal-Bench v2.1 for terminal agents, its 69.3 score surpasses Mistral Medium 3.5 by 18.7 points.
Pricing Strategy and the Open-Source Question
Quasar 438B is available via the CompactifAI API, priced at $0.60 per million input tokens** and **$1.80 per million output tokens. However, it is a proprietary model with undisclosed weights, supporting only text input and output, not images.
A crucial context: Multiverse Computing's core expertise lies in model compression. Its CompactifAI platform claims to compress models by 80-95% with minimal accuracy loss, but the company has not disclosed Quasar's compression ratio or its foundational architecture. While an official X post hinted that the model might be based on Zhipu's GLM5.2, Multiverse's documentation classifies it as an "original model."
Quasar 438B's release is a milestone for European AI, but its real significance may not be the 43 score itself—rather, it demonstrates that Europe can build usable enterprise-grade models. Yet, when a 27-billion-parameter Chinese model matches its score, the 438-billion-parameter count reveals a gap in compute efficiency and architectural innovation. In the world of AI, the title "Europe's best" is both an acknowledgment and a mirror.