Multiverse Computing launches Quasar 438B reasoning model
What's the deal? On 2 September 2026, Multiverse Computing introduced Quasar 438B, its first large model and flagship reasoning model for enterprise-scale agents and coding. The model has 438 billion parameters, supports English and Spanish, and is available through the company's CompactifAI API, allowing teams to test it without standing up their own infrastructure.
Benchmark performance: According to the company's comparison using the Artificial AnalysisDealroom has a profile for this one. Try Dealroom → Intelligence Index v4.1.1, Quasar scored 43, the highest score among the European models evaluated. The composite combines nine evaluations, including GDPval-AA, τ³-Banking, Terminal-Bench, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience and AA-LCR. The score was 13 points above Mistral Medium 3.5 and five points above NVIDIA Nemotron 3 Ultra, despite Nemotron having 112 billion more parameters.
Speed and capabilities: Quasar produced a 500-token response, including reasoning time, in 15.3 seconds. Multiverse reports that it combines long-context reasoning with agentic coding and terminal work: its AA-LCR score was 75.0, while its Terminal-Bench v2.1 score was 69.3. The company positions the model for software-development agents, technical copilots, research systems and workflow automation that require planning, tool use, code execution and large-context handling.
Strategic context: The launch extends Multiverse's focus on efficient and sovereign AI. Its CompactifAI technology compresses large language models to reduce computing requirements while retaining comparable performance and accuracy, supporting deployment across cloud, on-premises and edge environments. Quasar brings that efficiency work into the 400B-plus parameter class, with further updates promised by the company.
Read more: Multiverse Computing · Artificial Analysis · Multiverse Computing on X