Multiverse Computing Launches Quasar 438B
Multiverse Computing launched Quasar 438B, a 438-billion-parameter bilingual (English/Spanish) reasoning model for enterprise agents and coding It achieved the highest Artificial Analysis Intelligence Index score (43) among European models tested, outperforming Mistral Medium 3.5 (30) and NVIDIA Nemotron 3 Ultra (38) The model generates 500 output tokens in 15.3 seconds, combining high reasoning performance with low latency critical for agentic workflows Quasar scored 75.0 on Long Context Reason
Analysis
TL;DR
- Multiverse Computing launched Quasar 438B, a 438-billion-parameter bilingual (English/Spanish) reasoning model for enterprise agents and coding
- It achieved the highest Artificial Analysis Intelligence Index score (43) among European models tested, outperforming Mistral Medium 3.5 (30) and NVIDIA Nemotron 3 Ultra (38)
- The model generates 500 output tokens in 15.3 seconds, combining high reasoning performance with low latency critical for agentic workflows
- Quasar scored 75.0 on Long Context Reasoning (matching Grok 4.6 high) and 69.3 on Terminal-Bench v2.1, demonstrating strong coding and document-heavy task capabilities
- Available through the CompactifAI API, marking a significant milestone for European sovereign AI competitiveness against US and Chinese models
Why It Matters
Quasar 438B demonstrates that European AI developers can produce models that rival leading US and Chinese offerings in both reasoning capability and inference speed — a critical combination for enterprise deployment. Its focus on agentic workflows, where latency compounds across dozens of model calls, addresses a practical bottleneck that many large models struggle with in real-world enterprise settings.
Technical Details
- Model scale and architecture: 438 billion parameters, designed by Multiverse Computing (a compressed AI model specialist), balancing scale for demanding reasoning tasks with latency optimization for interactive enterprise use
- Benchmark performance: Artificial Analysis Intelligence Index v4.1.1 score of 43; AA-LCR score of 75.0 (matching Grok 4.6 high, within 1 point of Claude Opus 5); Terminal-Bench v2.1 score of 69.3 (18.7 points ahead of Mistral Medium 3.5)
- Speed: 500 output tokens in 15.3 seconds including reasoning time; faster than Mistral Medium 3.5 (18.8 seconds) while delivering a 13-point higher Intelligence Index score
- Evaluation framework: The Intelligence Index is a weighted average across nine evaluations in four categories — Agents (GDPval-AA v2, τ³-Banking), Coding (Terminal-Bench v2.1, SciCode), Scientific Reasoning (Humanity's Last Exam, GPQA Diamond, CritPt), and General knowledge/long-context reasoning (AA-Omniscience, AA-LCR)
- Availability: Accessible via CompactifAI API; bilingual support in English and Spanish; planned applications include software engineering, operational automation, and document-heavy research
Industry Insight
- European sovereign AI is reaching a competitive inflection point — Quasar's performance suggests that regional developers can offer viable alternatives to US-dominated models, which has implications for data sovereignty and regulatory compliance in enterprise procurement
- The emphasis on agentic latency (15.3s for 500 tokens) signals a shift in model design priorities: for enterprise agents making dozens of calls per task, speed optimization is becoming as important as raw benchmark scores
- The bilingual English/Spanish support and API-first distribution model reflect a strategic focus on practical enterprise adoption over pure research benchmarks, suggesting the next competitive frontier is deployment accessibility rather than parameter count alone
Disclaimer: The above content is generated by AI and is for reference only.