Artificial Intelligence
Nadia-1 Small: structured decisions in a compact model
Meet the Artinos decision model: how it works, its results and the limitations that accompany the research.
Bernardo B. Perobeli · Artinos
5 min read
Artinos introduces Nadia-1 Small: a multilingual model with 140.8 million parameters, adapted to choose among explicit alternatives. The research explores structured decisions with a compact model that can run on a CPU. It is not a free-form chatbot: its role is to assess options within the supplied context.
One interface for different decisions
Inputs combine context, an instruction, a decision type and alternatives. Outputs can be a categorical choice, a Boolean decision or an ordinal score. The model also returns the probability distribution, the selected index and the token count.
Consider a request: “I was charged twice for the same order.” With Billing, Delivery and Account access as options, the model can be evaluated for support routing. In another step, the options might represent priority levels. This illustrates a possible use, not a measured result for that request.
How Nadia-1 Small was built
The starting point is mmBERT-small, a pretrained multilingual encoder. Artinos adapted it with a shared head that scores alternatives, supervised training, and public and synthetic examples. The final stage combines textual inference and tool-use tasks with replay supervised by an earlier checkpoint.
The implemented limit is 2,048 tokens in total, including context, instructions and options. Larger inputs are rejected rather than silently truncated. Menu size and description length therefore matter when designing an integration.
What the tests showed
These results describe the selected Small checkpoint under the protocols documented in the study. Each evaluation measures a different task; the percentages do not constitute an overall model score.
SIB: 76.15% macro accuracy across 1,610 inputs in ten languages, with document-overlap controls against the local DEV set.
xSID: 58.63% macro accuracy by locale, across 8,750 inputs in 18 locales.
Typed: 40.40% accuracy across 2,000 decisions. Nadia-1 Small scored below Julia-1 under this protocol.
The model improves on some tasks and regresses on others. Some evaluation sets influenced development, and the complete history of test exposure is unknown. The study does not report an overall Decision Index ranking.
CPU execution, with measurement context
On the short English workload, Nadia-1 Small recorded a median latency of 47.61 ms and a P95 of 52.56 ms with eight threads. In the same measurement, Julia-1 recorded a 54.67 ms median and the original English version of Laya recorded 227.31 ms.
The test used 256 inputs, three repetitions, and a CPU environment provisioned with eight cores and 24 GiB of memory. Timings include the decision pipeline with the model already loaded, but exclude networking, queues, initialization and loading. They are not a latency guarantee for every computer or application.
Limitations that belong with the model
The order, names and descriptions of alternatives can affect decision quality. Returned probabilities should not automatically be treated as calibrated confidence. In additional evaluations, Small scored 12.33% on MMLU-Pro and 25.33% full-denominator accuracy on When2Call MCQ, exposing limits to generalization.
Deployment requires testing the actual questions, options, languages and error costs of an operation. High-impact decisions require human review. The study also documents unresolved data provenance and licensing issues; it does not establish the legal suitability of commercial distribution of the weights.
Research at Artinos
Nadia-1 Small investigates efficient decisions under resource constraints. The aim is to make both promising results and remaining weaknesses visible, without confusing a specialized model with general intelligence.
Source: “Nadia-1: adaptação econômica de um encoder multilíngue para decisões estruturadas”, Bernardo B. Perobeli, Artinos, October 2026. This article covers only the Small version and summarizes the manuscript supplied by its author.