Mio reduces AI model request costs 73.7%-97.5% in tests

SinapisAI Team said in an arXiv preprint published Oct. 20 that Mio 1.0 Lite delivered 73.7%-97.5% lower per-request costs than the best-scoring alternatives across nine public benchmarks.

The Oct. 20 arXiv preprint by SinapisAI Team frames the system as a way to coordinate multiple AI models under cost and resource limits, with a shared cache accelerating model-file retrieval by up to 23.3 times.

The paper, which is not yet peer reviewed, presents three components for what the authors call Collective Model Intelligence. Model Eval builds evidence-based capability profiles by combining query-level performance estimation with cross-domain prediction, reducing the need to test every model exhaustively.

Mio, short for Models in One, organizes complementary model candidates and supports task-conditioned, cost-aware selection as requirements change. Mos, or Model Serving OS, separates model selection from execution and handles endpoint scheduling, shared model deployment and distribution.

The authors said experiments showed lower capability-estimation errors when evaluation data were limited. They also said Mio 1.0 Lite achieved the highest average score and lowest average cost among the compared methods across the nine public benchmarks.

The work targets a growing operational problem in AI: organizations can access many specialized models, but using them together requires knowing which model fits a task, how to combine them and how to control serving costs. The preprint argues that simply adding more models does not itself produce Collective AI.

The limits are also clear. The paper runs 35 pages and includes 14 figures and eight tables.

Source: preprint, arXiv cs.AI, Oct. 20, 2026.