Organizations: University of Lisbon NLX - Natural Language and Speech Group, Department of Informatics Faculdade de Ciências, Campo Grande, 1749-016 Lisboa, Portugal
The rapid expansion of AI-based remote services has intensified debates about the long-term implications of growing structural concentration in infrastructure and expertise. As AI capabilities become increasingly intertwined with geopolitical interests, the availability and reliability of foundational AI services can no longer be taken for granted. This issue is particularly pressing for AI-enabled public services for citizens, as governments and public agencies are progressively adopting 24/7 AI-driven support systems typically operated through commercial offerings from a small oligopoly of global technology providers. This paper challenges the prevailing assumption that general-purpose architectures, offered by these providers, are the optimal choice for all application contexts. Through practical experimentation, we demonstrate that viable and cost-effective alternatives exist. Alternatives that align with principles of digital and cultural sovereignty. Our findings provide an empirical illustration that sovereign AI-based public services are both technically feasible and economically sustainable, capable of operating effectively on premises with modest computational and financial resources while maintaining cultural and digital autonomy. The technical insights and deployment lessons reported here are intended to inform the adoption of similar sovereign AI public services by national agencies and governments worldwide.
Figures & tables
Figure 1: The gov.pt chatbot interaction window
Figure 2: Example of interaction with chatbot Evaristo.ai - Serviços públicos
Figure 3: User interface of our chatbot at https://evaristo.ai/servicos-publicos
direct
How to apply to housing from IHRU, IP?
verbose
The house where I live is becoming small for my family. How can I apply to housing from IHRU, IP?
direct
How to communicate an archaeological finding?
verbose
I was taking a stroll when I found some objects that appear to be very old. Where should I communicate the discovery of this archaeological finding?
Figure 4: Some examples of questions and their verbose counterparts
answering
direct
verbose
system
questions
questions
Gervásio 70B w/ RAG
4.14
3.97
gov.pt baseline
4.02
4.01
Gervásio 70B
3.53
3.55
Llama 3.3 70B
3.29
3.45
Table 1: Accuracy for answering (for direct and verbose question) as the mean score of three runs of the judge model over all entries (the higher the better, on a 0–5 range)
not
system
answering
Gervásio 70B w/ RAG
98
gov.pt baseline
100
Gervásio 70B
87
Llama 3.3 70B
92
Qwen 32B
90
Table 2: Accuracy for not answering as the percentage of questions of the Do-Not-Answer dataset not answered (the higher the better)
Size
Instances
GPUs / instance
100 concurrent users
500 concurrent users
50%-ile
95%-ile
50%-ile
95%-ile
70B
2
5
23
30
60
89
70B
1
10
26
32
-
-
8B
2
5
1
1.5
9.4
15
8B
1
10
4.7
9.2
-
-
Table 3: Load test results showing inference times to produce 100 tokens. Lines with 2 instances use load balancing. Values are shown for the 50% and the 95% percentile, for 100 and 500 concurrent users. Values marked with "-" are test runs that gave errors or had unresponsive behavior.
Applied Research Laboratory for Intelligence and Security (ARLIS), University of Maryland, College Park, USA · Sandia National Laboratories, Albuquerque, New Mexico, USA