Conceptualising local AI infrastructure using small language models (SLMs) for primary education

1 Introduction

The integration of artificial intelligence (AI) into educational contexts has accelerated dramatically in recent years (Chen et al., 2020; Pedro et al., 2019). Large language models (LLMs) such as ChatGPT have captured widespread attention, demonstrating capabilities that include content generation, explanation of complex concepts, and adaptive questioning (Annepaka and Pakray, 2025; Hadi et al., 2023). For educators and researchers, such systems appear to offer remarkable potential for personalised learning support, assessment scaffolding, and pedagogical innovation (Holmes, 2020; Nwachukwu et al., 2025; Selwyn, 2022).

However, the deployment of cloud-based LLMs in primary schools encounters formidable barriers. Chief among these are concerns regarding the privacy and security of child data (Salah et al., 2026; Tsouplaki et al., 2025). Primary schools serve populations whose data protection is governed by increasingly stringent regulatory frameworks, particularly the Children’s Online Privacy Protection Act in North America (Steeves and Mačėnaitė, 2022) and similar data protection regulations in the European Union and Commonwealth nations (Livingstone and Helsper, 2008). Beyond regulatory compliance, primary schools also face substantial practical constraints: limited information technology budgets, inconsistent broadband infrastructure, particularly in rural areas, and the absence of specialist AI expertise within school leadership and teaching staff (Kormos and Wisdom, 2023; Waiganjo and Mutemaringa, 2025).

The reliance on cloud-based AI systems also introduces what might be termed a ‘surveillance risk’ at the institutional level. When student interactions are processed through cloud-based platforms managed by private technology companies, educational institutions may lose substantial control over the collection, retention, and subsequent use of student data. This situation raises significant concerns regarding privacy, data governance, and accountability, particularly in light of growing evidence that some educational technology providers engage in data monetisation practices, whereby student information is leveraged for commercial purposes beyond its original educational intent (Zeide, 2017).

This paper proposes a fundamentally different architecture for educational AI in primary schools: the deployment of small language models (SLMs) (Wang et al., 2025; Zhang et al., 2025) on school-owned hardware, accessible exclusively through school networks, with all student data remaining within institutional boundaries. SLMs, such as Microsoft Phi (Abdin et al., 2024), Google Gemma (Team et al., 2024), Meta Llama small variants (Abdullah et al., 2025), and Mistral models (Karamcheti et al., 2021), have recently emerged as computationally efficient alternatives to frontier LLMs. These systems require significantly fewer computational resources, can operate offline, and offer substantially lower inference costs whilst maintaining performance on many educationally relevant tasks (Wang et al., 2025; Zheng et al., 2025). Our thesis is that locally hosted SLMs, operating within school network infrastructure, may offer a feasible middle ground between educational innovation and robust child data protection. This is not an argument for technological determinism, but rather a conceptual framework for understanding how particular institutional contexts shape which AI architectures are most pedagogically and ethically appropriate.

Although empirical studies examining the classroom use of SLMs with primary school students remain limited, emerging research from computational linguistics provides relevant insights into their potential alignment with child language development processes. Recent studies have explored how SLMs generate texts that resemble those produced by young learners, highlighting structural and linguistic similarities in writing patterns (Buhnila et al., 2025). Other work has investigated the ability of small models to simulate developmental linguistic phenomena such as overregularisation errors (e.g., “goed” instead of “went”), which are characteristic of early language acquisition (Haga et al., 2024). Additionally, research on small-scale models in multilingual settings suggests that SLMs can approximate aspects of bilingual language acquisition and developmental learning trajectories (Zeng et al., 2026). While these studies do not evaluate educational outcomes or classroom effectiveness, they indicate that SLMs can capture features of child-like language production and learning processes. This emerging body of work provides a useful theoretical foundation for considering how SLMs might be adapted for educational purposes, while also underscoring the current lack of empirical validation in real school settings. However, these findings should be interpreted cautiously, as they are based on computational simulations and textual analyses rather than direct classroom implementation with children.

It is important to emphasise that this paper is conceptual in nature and does not report findings from a deployed educational AI system. Rather than evaluating learning outcomes or system performance, the paper develops a theoretically informed framework that integrates technological, educational, and governance perspectives to examine the institutional suitability of locally hosted SLMs in primary education. Accordingly, the arguments presented should be interpreted as conceptual propositions regarding potential benefits, limitations, and implementation pathways. Empirical validation through pilot deployments, classroom studies, and stakeholder evaluations remains necessary before conclusions regarding effectiveness can be established.

The paper proceeds as follows. We first define SLMs and their technical characteristics and compares them with cloud-based LLMs. We then examine why primary schools represent a distinctive and uniquely suitable context for this architecture. We present a concrete infrastructural model comprising hardware, software, and network architecture. We analyse the privacy and governance advantages of on-premises deployment. We discuss realistic educational use cases, acknowledge limitations and risks, and finally situate this approach within broader theoretical frameworks of digital sovereignty and AI governance in education.

2 What are small language models?

SLMs are compact neural network architectures designed to perform natural language processing (NLP) tasks with substantially reduced computational requirements compared to frontier LLMs (Wang et al., 2025). Whilst frontier models such as GPT-4 contain hundreds of billions of parameters, SLMs typically contain between 1B and 13B parameters. Notable examples include Microsoft Phi-3 (3B–14B parameters), Google Gemma (2B–7B parameters), Meta Llama 2 small variants (7B parameters), and Mistral 7B.

These models are distinguished by several technical and operational characteristics. First, they require substantially lower computational resources: SLMs can execute effectively on consumer-grade graphics processing units (GPUs) or even central processing units (CPUs) with appropriate optimisation techniques. Second, they can operate offline, requiring no external network connectivity beyond initial model download. Third, they offer dramatically reduced inference costs compared with cloud-based API services. Fourth, they demonstrate lower energy consumption, an important consideration for schools seeking to minimise operational carbon footprints. Fifth, they permit rapid institutional deployment without vendor lock-in. Sixth, they enable meaningful governance, in that schools retain direct control over model parameters, fine-tuning, and access logs (Abdin et al., 2024; Wang et al., 2025; Zhang et al., 2025).

However, SMLs also exhibit genuine limitations. They typically demonstrate reduced reasoning depth compared with frontier models, particularly on tasks requiring multi-step logical inference or specialised domain knowledge (Fu et al., 2023). They have smaller context windows, limiting the quantity of preceding text they can reference when generating responses. Their general capability across diverse domains is narrower than that of larger frontier models. They demonstrate higher hallucination rates, defined as the generation of plausible sounding but factually incorrect information (Kalai et al., 2025). Their knowledge is constrained by training data, and they perform less effectively on out-of-distribution tasks.

Despite these limitations, we propose that SLMs are well-suited to primary educational contexts precisely because primary school learning tasks are frequently characterised by particular properties. Such tasks are often domain-specific, focused on defined curriculum areas. They are repetitive and structured, with consistent formats and expected response patterns. They are heavily scaffolded, with teacher guidance structuring student interactions. They are age-appropriate, with simplified language and conceptual frameworks. They require human mediation, in that teachers retain pedagogical authority and evaluate model outputs before student exposure. Under these conditions, the limitations of SLMs become far less problematic, whilst their advantages in governance, privacy, cost, and deployability become highly significant.

2.1 Comparing SLMs and cloud-based LLMs

While SLMs provide advantages in deployment flexibility, governance control, and infrastructure independence, cloud-based LLMs continue to offer substantial benefits that should not be overlooked. Frontier models such as GPT-4, Claude, and Gemini generally demonstrate superior reasoning capabilities, broader knowledge coverage, larger context windows, stronger multilingual performance, and more robust handling of complex instructional tasks (Annepaka and Pakray, 2025; Meva and Kukadiya, 2025). These capabilities may be particularly valuable in educational scenarios requiring advanced explanation generation, adaptive tutoring, or complex problem-solving support.

Cloud-based deployment also transfers substantial operational responsibilities from schools to technology providers. Infrastructure maintenance, security patching, hardware upgrades, model optimisation, and disaster recovery processes are typically managed by vendors with significant technical expertise and resources. This may reduce implementation complexity for schools that lack dedicated information technology personnel (Selwyn, 2022).

However, these benefits are accompanied by trade-offs relating to data governance, institutional autonomy, recurring subscription costs, and dependency on external providers. Consequently, the choice between local SLM deployment and cloud-based LLM services should not be understood as a binary distinction between superior and inferior technologies. Rather, it represents a contextual decision involving competing priorities including performance, privacy, governance, sustainability, and organisational capacity. Different educational institutions may reasonably arrive at different conclusions depending on their specific circumstances.

Table 1 illustrates that the choice between locally hosted SLMs and cloud-based LLMs involves a series of trade-offs rather than a clear technological superiority of one approach over the other. Cloud-based systems generally provide stronger technical performance and reduced infrastructure management requirements, whereas locally hosted SLMs may offer advantages relating to institutional control, data governance, and operational independence. Consequently, the appropriateness of either approach should be evaluated in relation to the specific educational, regulatory, financial, and organisational circumstances of the school rather than solely on the basis of model capability.

TABLE 1. Comparative characteristics of locally hosted SLMs and cloud-based LLMs in primary education contexts (yellow, limitation; green, advantage; blue, balanced).

3 Why primary schools are a unique context

Primary schools, serving children typically aged five to 12 years, represent a distinctive institutional and regulatory context that shapes appropriate technological approaches to AI integration (Yang, 2022). This uniqueness operates across developmental, institutional, and regulatory dimensions.

3.1 Developmental considerations

Children in primary school years are developing cognitive capabilities including executive function, abstract reasoning, and metacognitive awareness (García et al., 2016; Piaget, 2013). They require adult mediation to distinguish reliable from unreliable information sources and to recognise when an AI system has produced incorrect or misleading outputs. Children are developmentally vulnerable users who may be disproportionately affected by exposure to inappropriate AI-generated content and by interaction patterns that encourage excessive or potentially addictive engagement. These concerns underscore the need for age-appropriate safeguards, content moderation, and ethical design practices in educational AI systems (Livingstone, 2008). They require simplified language and conceptual scaffolding. They benefit from predictable, consistent interactions, particularly with systems designed to support rather than replace human pedagogical relationships. These developmental factors necessitate AI systems that are transparent, easily explicable to young learners, and thoroughly mediated by educators.

3.2 Institutional constraints

Primary schools in many jurisdictions, particularly those in rural areas or economically disadvantaged regions, face substantial resource constraints. Information technology budgets are typically limited, often allocated first to essential infrastructure such as email systems, learning management platforms, and connectivity (Hwang et al., 2020; Pedro et al., 2019). Technical support staffing is minimal, frequently consisting of a single part-time information technology technician responsible for an entire school. Broadband infrastructure is inconsistent, with rural schools particularly affected by unreliable or low-bandwidth connections. Specialist expertise in AI and emerging technologies is rare among school leadership and teaching staff. Schools frequently lack the purchasing power to enter long-term vendor contracts for expensive cloud-based AI platforms.

These constraints mean that technological solutions appropriate for well-resourced institutions may be entirely unsuitable for the primary school context. Cost escalates rapidly at scale; a primary school with 400 students cannot absorb the cumulative cost of cloud-based API calls across multiple classrooms. Dependence on external cloud infrastructure introduces failure modes that schools cannot mitigate: loss of internet connectivity during storms or infrastructure disruptions, unexpected price increases by vendors, or suspension of service due to changes in corporate policy. The absence of local technical expertise means schools must rely entirely on external vendor support for troubleshooting.

3.3 Regulatory and governance landscape

Data protection frameworks have become increasingly stringent with respect to child data. The European Union’s General Data Protection Regulation (Das, 2018) and its specific provisions for data concerning children establish that processing of child data requires heightened legal safeguards. Numerous jurisdictions have enacted specific protections for children’s online data, including age-appropriate privacy controls and restrictions on data retention and secondary uses (Livingstone and Helsper, 2008). Parental concerns regarding surveillance and data commodification in educational technology have grown substantially, contributing to increased scrutiny of school technology adoption decisions (Boyd, 2014).

When schools deploy cloud-based AI systems, they transfer responsibility for compliance with these regulations to the cloud vendor, yet they retain ultimate liability for data breaches or regulatory violations. When schools cannot fully understand vendor data handling practices, audit their compliance, or enforce contractual obligations, they accept substantial governance risk. Educational AI in primary settings is, fundamentally, not merely a technical challenge but a governance and safeguarding challenge. The question is not solely whether an AI system is intelligent, but whether school leadership can meaningfully explain to parents and regulators what is happening to their children’s learning data.

4 Conceptualising an in-school small language model server

We conceptualise an on-premises SLM infrastructure as a specific sociotechnical configuration in which a school deploys one or more discrete local servers hosting SLMs, accessible exclusively through the school’s internal network, with explicit architectural prohibitions on external data transmission. This configuration differs fundamentally from cloud-based architectures, in which schools are customers of external computational infrastructure, subject to vendor terms of service and reliant on network connectivity to external servers. The on-premises model has the potential to provide better computational capacity, data governance, and network control to the school institution itself. Figure 1 shows the visual illustration of an example setup to understand the concept.

4.1 Hardware architecture

We propose that a primary school might deploy one or more dedicated servers hosting SLMs. These servers need not be expensive specialist hardware. A cost-effective implementation might utilise a refurbished desktop workstation with a mid-range GPU unit, or a mini PC equipped with an entry-level GPU suitable for inference operations. Depending on hardware specifications and local procurement conditions, implementation costs may range from several hundred to several thousand Australian dollars per server, representing a one-off capital expenditure rather than ongoing subscription costs.

Alternatively, schools might repurpose existing network storage infrastructure, deploying models on network-attached storage devices equipped with sufficient computational capacity. A single server could potentially host multiple SLM instances, distributed across different classroom networks or accessible through a unified interface, depending on institutional topology and requirements.

4.2 Software architecture

The software stack for on-premises model deployment has matured substantially. We propose that schools might utilise established open-source frameworks such as Ollama (Marcondes et al., 2025), a simplified interface for deploying and managing open-source language models, or LM Studio, a desktop application providing graphical interface and local API endpoints for model inference. These frameworks will potentially eliminate the need for schools to engage with complex low-level deep learning libraries.

Model selection can be tailored to educational purposes. For literacy support and language-based tasks, Google Gemma or Microsoft Phi models may offer strong performance. For reasoning tasks and curriculum explanation, Mistral or Llama 2 models might be appropriate. All suggested models are open-source, hosted on repositories such as Hugging Face, and available for download without licensing restrictions.

Schools would interact with these models through a web-based interface, potentially a simplified dashboard providing teachers and students with access to model inputs and outputs. Such interfaces might provide basic logging and transcript retention for transparency and safeguarding audit purposes.

4.3 Network architecture

The network architecture must ensure that all student interaction occurs exclusively within school-controlled networks. Student devices (e.g., tablets, laptops, classroom computers) connect to SLM server exclusively through the school’s internal local area network (LAN) or intranet. No student interaction data is transmitted externally. The server itself should be configured without outbound internet connectivity, preventing accidental or intentional data exfiltration.

Schools might implement a basic teacher dashboard, accessible through secure school-network login, providing educators with visibility of interaction logs, usage patterns, and student engagement. This transparency will potentially serve both pedagogical and safeguarding purposes, enabling teachers to monitor interactions and respond to concerning content or engagement patterns.

5 Privacy and data governance advantages

The deployment of SLMs on school-owned, locally controlled infrastructure may offer substantial privacy and governance advantages compared with cloud-based AI systems. However, these advantages should not be interpreted as absolute. Contemporary cloud providers increasingly implement sophisticated security controls, privacy safeguards, and compliance mechanisms. The key distinction lies not in whether privacy protections exist, but in who retains authority over data governance decisions and infrastructure management.

• Data Sovereignty: When students interact with a school-hosted SLM, their input prompts and the model’s responses remain entirely within school infrastructure. No student data is transmitted to external cloud servers operated by private corporations. This may support stronger forms of institutional data sovereignty: the school institution retains exclusive access to records of student learning interactions. While cloud-based providers increasingly implement privacy controls and enterprise governance mechanisms, schools remain dependent on vendor-specific policies and infrastructure. For example, ChatGPT, in which student prompts are transmitted to OpenAI servers, remain accessible to OpenAI’s staff, and are potentially subject to future data monetisation or secondary use.
• Reduced Third-Party Dependence: Cloud-based AI services create dependence on third-party vendors for the security, availability, and ethical appropriateness of AI systems. Schools cannot audit whether OpenAI’s data security practices meet regulatory requirements. Schools cannot enforce contractual obligations to delete student data. Schools cannot determine how long prompts are retained on OpenAI’s servers or what secondary purposes they are used for. By contrast, on-premises deployment may eliminate this third-party dependence. Schools own the infrastructure, the models, and the data. Teachers will be able to verify that the server is not transmitting data externally. School leadership potentially can design and enforce data retention policies directly.
• Regulatory Compliance Pathways: On-premises deployment substantially may simplify regulatory compliance. Schools will be able to implement data retention policies that automatically delete interaction logs after specified periods, ensuring GDPR and similar regulations are honoured. Schools may demonstrate to parents, regulators, and school boards precisely what data is collected, where it is stored, and how it is protected. Schools can potentially audit access logs to verify that only authorised staff members access student data. They can provide detailed explanations of the AI system’s training data, capabilities, and limitations. This transparency is difficult or impossible with commercial cloud-based systems, where training data, model architecture, and data handling practices are proprietary and inaccessible to schools.
• Offline Capability and Connectivity Resilience: An important likely practical advantage of on-premises SLMs is their offline functionality. A school-hosted model continues to operate during internet outages, may provide educational value even in environments with unreliable broadband connectivity. This will be particularly significant for primary schools in rural or remote areas, where connectivity may be intermittent. It may also educationally valuable during entire-school activities such as excursions, where portable devices with downloaded models could still access AI-supported learning resources. This resilience is impossible with cloud-dependent systems, which cease functioning entirely when internet connectivity is lost (Brahimi and Sarirete, 2015).
• Balancing Privacy and Risk: We acknowledge that on-premises deployment does not eliminate all privacy and security risks. Schools remain responsible for physical security of their servers, for cybersecurity practices including access controls and encryption, and for staff training to prevent accidental data exposure. Malicious actors could potentially compromise poorly secured school servers. Staff misconduct could violate data protection policies. These risks are real and require schools to implement appropriate safeguarding measures. However, local hosting shifts privacy governance from external commercial entities toward school controlled institutions. Schools can implement data minimisation practices, conduct security audits, and enforce accountability structures in ways that are impossible with cloud-based systems operated by private technology corporations.

6 Educational use cases in primary schools

SLMs, appropriately designed for on-premises deployment, may support numerous educationally valuable functions within primary school contexts. These use-cases are intended as illustrative examples of potential applications rather than empirically verified educational outcomes. Their effectiveness would depend on implementation quality, teacher mediation, curriculum alignment, and institutional context.

• Literacy Support: SLMs can potentially support literacy instruction through generating vocabulary explanations, creating reading comprehension prompts tailored to specific texts, providing sentence starters for student writing, and offering simplified explanations of complex literary concepts. A student reading a text at their current instructional level could ask the model to explain difficult vocabulary words. A teacher preparing literacy lessons could use the model to generate discussion questions aligned to specific learning objectives. These applications are well within the capabilities of SLMs, which perform strongly on tasks involving text manipulation and explanation.
• Teacher Support and Curriculum Assistance: Teachers can potentially utilise SLMs as support tools for curriculum preparation. Models can generate differentiated worksheet content for students at different achievement levels, suggest strategies for scaffolding particular concepts, draft rubrics for assessment, provide curriculum explanations across subject areas, and brainstorm pedagogical strategies for presenting difficult content. A teacher preparing a unit on the water cycle could request simplified explanations suitable for younger students and more complex versions for advanced learners. These applications reduce teacher administrative burden and support pedagogical personalisation.
• Classroom Inquiry and Guided Learning: SLMs has the potential to participate in guided question-and-answer interactions with students, supporting structured classroom inquiry activities. For example, during a unit on solar systems, a teacher might structure an activity in which students ask the model questions about planetary properties, with the teacher providing guidance and correction. The model could respond to student curiosity with age-appropriate explanations. These guided interactions scaffold student questioning and reasoning skills whilst maintaining teacher pedagogical authority. The teacher always mediates and evaluates the model’s responses before students encounter them.
• Inclusion and Personalisation: For students with learning difficulties, language disorders, or English as an additional language, SLMs can potentially generate simplified content explanations, multilingual support, and personalised pacing. A student with a reading difficulty could receive curriculum explanations in simplified language. A student learning English as an additional language could access vocabulary scaffolding. A student with attention difficulties might interact with focused, structured questions rather than dense textbook content. These personalisation capabilities support inclusive education, though always under teacher supervision and direction.

7 Risks and limitations

We emphasise that deployment of SLMs in primary schools, even on-premises, introduces genuine risks and limitations that require careful institutional management.

• Hallucination and Factual Unreliability: SLMs are known to generate plausible sounding but factually incorrect information, a phenomenon termed hallucination. A model might confidently state that the capital of Australia is Melbourne or provide false historical information. For young students still developing critical evaluation skills, such hallucinations present an educational risk. Students may incorrectly absorb false information, particularly when generated by a system framed as an educational tool. Schools must implement safeguarding strategies: teachers must always review model outputs before student exposure, and students must be taught to treat AI-generated information as provisional rather than authoritative.
• Bias and Representation: Language models learn statistical patterns from training data, and when training data reflects historical biases, models can perpetuate and amplify those biases. A model might generate content that stereotypes particular social groups, overrepresents certain perspectives whilst marginalising others, or reflects cultural assumptions embedded in its training data. Teachers must be prepared to identify and correct biased outputs, and schools should monitor model behaviour for systematic patterns of bias. This is an institutional responsibility, not something that technology alone can address (Bolukbasi et al., 2016; Buolamwini and Gebru, 2018).
• Technical Maintenance Burden: Deploying on-premises infrastructure requires school staff to manage system maintenance: hardware troubleshooting, software updates, model updates, and addressing technical failures. Primary schools typically lack specialists with the technical expertise to independently manage these tasks. Schools may need to hire external consultants or contractors to manage on-premises infrastructure, introducing costs not present in cloud-based models. System failures may lead to extended periods when the AI system is unavailable, disrupting planned learning activities.
• Teacher Digital Literacy Requirements: Teachers must possess adequate understanding of SLM capabilities and limitations in order to use them effectively and safely. Teachers need to understand what kinds of questions the models can reliably answer, when models are likely to hallucinate, how to recognise biased outputs, and how to frame AI interactions as educationally scaffolded rather than autonomous. This requires professional development and ongoing support. Not all teachers will feel equally confident with AI technology, and schools must provide training and support rather than assuming teachers will independently develop necessary competencies (Holmes, 2020).
• Safeguarding and Content Moderation: Schools must implement safeguarding policies that define appropriate use of AI systems and establish mechanisms for responding to concerning interactions. SLMs may generate inappropriate content if prompted with offensive or inappropriate input. Schools need content filtering policies and processes for addressing instances in which students access unsuitable material. This requires ongoing institutional oversight, not merely technical implementation.
• Uneven Output Quality: Model output quality is not uniform. Some student queries will receive excellent, pedagogically useful responses. Others will receive incomplete, confusing, or incorrect answers. This inconsistency can be frustrating for students and teachers and may undermine confidence in the system. Institutional readiness requires understanding that SLMs are imperfect tools that augment rather than replace teacher expertise.
• Cybersecurity and Infrastructure Security: While local deployment reduces dependence on external service providers and may strengthen institutional control over student data, it simultaneously transfers cybersecurity responsibility to the educational institution. Schools become responsible for protecting infrastructure against unauthorised access, malware infections, ransomware attacks, network intrusions, insider threats, and accidental data exposure. This responsibility is particularly significant given the increasing frequency of cyberattacks targeting educational institutions and the sensitivity of student information maintained within school systems. Effective implementation therefore requires comprehensive cybersecurity governance, including role-based access controls, multi-factor authentication, encryption of stored data, network segmentation, secure backup and recovery procedures, vulnerability assessments, regular software patching, and ongoing security monitoring. Schools must also establish incident response and disaster recovery plans to minimise disruption and ensure compliance with applicable privacy and data protection regulations. Consequently, local deployment should not be viewed as eliminating privacy and security risks; rather, it redistributes responsibility from external vendors to educational institutions, making organisational cybersecurity capability and governance maturity critical determinants of successful implementation.
• Scalability and Infrastructure Growth: Scalability represents an important practical consideration that is frequently overlooked in discussions of local AI deployment. While a single SLM server may adequately support a classroom-scale implementation or pilot project, larger schools may require multiple servers, additional storage capacity, distributed computing resources, or load-balancing mechanisms to accommodate simultaneous users and increasing demand. As adoption expands across year levels, subject areas, and administrative functions, processing requirements, network traffic, and infrastructure complexity may increase substantially. Furthermore, future generations of language models may require greater computational resources than those available within existing school infrastructure, necessitating periodic hardware upgrades and system redesign. Educational institutions considering long-term implementation must therefore evaluate not only current requirements but also projected growth in user demand, model complexity, and technological expectations. While local deployment may initially appear economically attractive and operationally manageable, scalability challenges may influence its long-term sustainability and require additional investment in infrastructure and technical expertise.
• Total Cost of Ownership: Comparisons of cost-effectiveness between locally hosted SLMs and cloud-based AI services should extend beyond initial acquisition or subscription costs and instead consider the broader concept of total cost of ownership (TCO). Although local deployment may reduce recurring licensing and API expenses, schools must also account for electricity consumption, hardware replacement cycles, software maintenance, cybersecurity management, technical support contracts, staff training, professional development programs, infrastructure upgrades, and the potential costs associated with system downtime or service interruptions. Conversely, cloud-based solutions often provide predictable subscription-based pricing while reducing internal infrastructure and maintenance requirements. The relative economic advantages of either approach are therefore likely to vary according to factors such as school size, implementation scale, usage intensity, available technical expertise, and deployment duration. At present, no comprehensive empirical studies exist comparing lifecycle costs of locally hosted SLMs and cloud-based LLMs in primary education settings. Therefore, the discussion presented here should be interpreted as a conceptual framework for future economic evaluation rather than a definitive cost comparison.

8 Educational foundations of local SLM adoption

8.1 Constructivist perspectives

Constructivist theories of learning argue that knowledge is actively constructed by learners through interaction with their environment rather than passively transmitted from teacher to student (Piaget, 2013). From this perspective, AI systems should not be viewed as replacements for instruction but as tools that support exploration, reflection, and knowledge construction. Locally hosted SLMs may function as interactive cognitive resources that enable students to ask questions, explore concepts, and receive explanatory support tailored to their current level of understanding. Such applications align with broader views of educational technology as a means of enhancing learner agency rather than automating teaching processes (Papert, 2020). The value of AI therefore lies not primarily in generating answers but in facilitating opportunities for inquiry, dialogue, and conceptual development.

8.2 Social constructivism and scaffolding

Social constructivist theories emphasise the role of guided interaction in supporting learning. Vygotsky’s concept of the Zone of Proximal Development (ZPD) suggests that learners can achieve higher levels of performance when supported by knowledgeable others (Vygotsky, 1978). Bruner (1997) further developed the concept of scaffolding to describe temporary support structures that assist learners in accomplishing tasks beyond their independent capabilities. Within this framework, SLMs may function as supplementary scaffolding tools that provide explanations, prompts, examples, and guided questioning. However, the role of the teacher remains central. Educators must determine when AI-generated assistance is appropriate, evaluate the accuracy of outputs, and ensure that learning activities remain aligned with curriculum objectives. Consequently, local SLMs should be understood as augmenting rather than replacing teacher expertise.

8.3 Classroom implementation challenges

Translating conceptual potential into classroom practice presents significant challenges. Teachers must develop sufficient understanding of AI capabilities and limitations to evaluate outputs critically and integrate them appropriately into instructional activities. Professional development therefore becomes a prerequisite for successful implementation (Holmes, 2020). Additional challenges include curriculum alignment, assessment integration, classroom management considerations, safeguarding requirements, and ensuring equitable access among diverse student populations. Students must also develop AI literacy competencies that enable them to question, evaluate, and critically interpret AI-generated information rather than accepting outputs uncritically (Tang et al., 2025).

9 Conceptualising educational AI through institutional appropriateness

Educational technology research has traditionally framed technological innovation as inherently progressive, evaluating systems primarily by their technical sophistication and performance capabilities. Within this framework, a LLM like GPT-4 is unambiguously superior to a smaller model like Phi-3, on the basis of greater reasoning capability and broader knowledge. However, we propose an alternative conceptual framework, grounded in institutional context and educational governance.

Within this framework, the educational value of an AI system is not determined solely by its raw intelligence. Rather, it is determined by a constellation of factors including institutional capacity to deploy and maintain the system, institutional authority to govern student data and interactions, pedagogical fit for particular educational contexts, cost and sustainability, regulatory compliance, and alignment with schools’ capacities and values. From this perspective, a SLM hosted locally on school infrastructure may be profoundly more educationally valuable than a sophisticated Frontier model whose operation schools cannot audit, whose data practices they cannot control, and whose costs they cannot sustain.

This argument resonates with broader theoretical frameworks in critical technology studies and educational technology research. Selwyn (2022) argues that technological determinism within educational contexts obscures the ways in which technology reflects and reinforces power relationships, corporate interests, and particular ideologies. Rather than asking ‘What can AI do for education?’, schools might more productively ask ‘What kind of AI aligns with our institutional values and governance capacity?'

The concept of digital sovereignty, increasingly prominent in policy discussions across Europe and elsewhere, emphasises institutional and national capacity to govern digital technologies in alignment with local values and contexts (Robles-Carrillo, 2023). Educational AI presents a case where local digital sovereignty becomes particularly important: when primary schools rely on external, proprietary technology platforms to deliver education to young people, schools sacrifice meaningful control over a dimension of childhood socialisation and development.

Privacy-by-design, a framework originating in data protection and design ethics, proposes that privacy protection should be embedded into technical systems from initial design, rather than added as an afterthought (Cavoukian, 2009). On-premises deployment of SLMs represents a privacy-by-design approach to educational AI: privacy protection is not a regulatory compliance burden, but an integral feature of how the system is architecturally organised.

We propose that in educational contexts, particularly primary education, ‘smaller but controllable’ may be institutionally more valuable than ‘larger but opaque’. This is not because smaller language models are objectively superior technologies. It is because primary schools are, fundamentally, institutions responsible for child protection and human development. School leadership and teaching staff need to be able to confidently explain to parents, governors, and regulators what is happening to children’s learning data and how AI systems serve pedagogical purposes. Institutional transparency, governance authority, and protective capacity matter profoundly. SLMs hosted on school infrastructure enable this institutional control in ways that cloud-dependent Frontier models do not.

10 Conclusion

SLMs are not replacements for Frontier AI systems. They have genuine technical limitations, including higher hallucination rates, smaller context windows, and narrower general capability. They are not appropriate for tasks requiring sophisticated reasoning or specialised domain expertise far beyond primary school contexts. However, we have argued that SLMs represent a theoretically grounded and potentially viable pathway for integrating AI into primary school contexts. Their lower cost, offline functionality, reduced computational requirements, and open-source availability make them accessible to schools that cannot sustain cloud-based subscriptions. More importantly, their on-premises deployment may enable schools to maintain institutional governance over student data and learning interactions. The propositions advanced throughout this paper should not be interpreted as validated outcomes. While the conceptual analysis suggests that locally hosted SLM infrastructures may offer promising pathways for balancing educational innovation with privacy protection and institutional governance, the practical effectiveness of such systems remains to be demonstrated.

The value of this architecture is not primarily technical. It is ethical and institutional. When primary schools deploy AI systems on infrastructure they control, with data remaining within their boundaries, they maintain the capacity to answer fundamental questions: What data are we collecting from children? How long are we retaining it? Who has access to it? What are we using it for? Are we treating child data as a resource to be stewarded, or as a commodity to be monetised? These questions are increasingly difficult to answer when schools rely on external proprietary platforms operated by corporations whose primary accountability is to shareholders rather than to children and families.

Future research should investigate multiple dimensions of local SLM deployment in educational settings. First, pilot studies should evaluate technical feasibility, reliability, and infrastructure requirements under real-world classroom conditions. Second, comparative studies should assess educational outcomes associated with local SLMs and cloud-based LLMs, including impacts on student engagement, learning achievement, teacher workload, and instructional quality. Third, governance-oriented research should investigate stakeholder perceptions, including those of teachers, parents, school leaders, and policymakers. Fourth, economic analyses should evaluate total cost of ownership across multi-year implementation periods. Finally, longitudinal studies are needed to examine how local AI deployment influences institutional trust, perceptions of data sovereignty, and long-term educational adoption patterns.

Policymakers should consider whether current educational technology procurement policies, which often incentivise large-scale cloud-based platforms, inadvertently push schools toward dependency on commercial vendors at the expense of institutional autonomy. Governments might support research, policy development, and funding mechanisms that enable and encourage schools to adopt locally hosted, open-source AI infrastructure.

Most importantly, we emphasise that the central question in educational AI is not ‘How intelligent can our systems be?’ but ‘What kind of systems enable schools to provide education that is protective of children, transparent to families, and answerable to democratic communities?’ A locally hosted SLM infrastructure offers a conceptual model for educational AI that prioritises institutional autonomy, child data protection, and equitable access over raw computational power. In a landscape increasingly dominated by commercial edtech platforms, this represents a modest but potentially significant alternative.

Ultimately, the contribution of this paper lies not in demonstrating the superiority of locally hosted SLMs over cloud-based alternatives, but in proposing a conceptual framework through which educational institutions may evaluate AI deployment choices. The framework suggests that institutional appropriateness, governance capacity, privacy protection, and educational context should be considered alongside technical performance when making decisions regarding educational AI adoption. Whether locally hosted SLMs ultimately deliver the anticipated educational, governance, and privacy benefits remains an empirical question that warrants future investigation.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. This work utilized Generative AI platforms, specifically GPT and Gemini, for text refinement, grammatical improvement, and the conceptual transformation of ideas into visual illustrations. The authors maintain full responsibility for the accuracy, integrity, and correctness of the final content.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Minxian Xu, Chinese Academy of Sciences (CAS), China

Reviewed by: Pramila Ramani, Central University of Tamil Nadu, India

Senad Orhani, University of Prishtina, Albania