1. Introduction
Open education has emerged as a strategic response to the need to transform education systems worldwide. It has become an essential response to the contemporary challenges of knowledge democratization and reduction of educational gaps (Landers & Behrend 2023). However, integrating emerging technologies like Open Large Language Models (OLLMs) poses a new challenge: ensuring that these high-impact, computationally demanding artificial intelligence tools can be adopted sustainably in globally diverse contexts (Aggarwal 2023; UNESCO 2021). This problem opens the door to new strategies combining technological innovation with educational equity principles.
Faced with this challenge, computational optimization of OLLMs is presented as a crucial strategy to close this gap. Techniques such as unstructured pruning and Retrieval-Augmented Generation (RAG) make it possible to significantly reduce processing and memory requirements without compromising the quality of educational assistance provided (Guo et al. 2023; Lewis et al. 2021). This approach improves operational efficiency and expands access to academic support technologies to low-income institutions, self-taught students, and technologically vulnerable communities (Ferrara 2023). Thus, implementing optimization processes is not only a matter of engineering but also an act of educational justice that demands priority attention in technological integration policies.
This study aimed to demonstrate the practical applicability of OLLM optimizations in open and globally diverse educational contexts. Optimizing OLLMs can enable equitable access to knowledge, supporting the principles of open education and science (Bandi, Adapa & Kuchi 2023; Dakakni & Safa 2023). An experimental analysis in five open-source models resulted in practical strategies to implement artificial intelligence solutions that respect the ideals of inclusion, sustainability, and transparency that will govern the future of global education (Camacho-Zuñiga et al. 2024; Luckin et al. 2022). This research, therefore, not only contributes to the academic debate but also offers practical tools to transform technological accessibility into concrete educational realities. The originality of this study lies in its integrative methodology that combines unstructured pruning, RAG, LoRA, and the concept of a living educational memory to enable scalable, adaptive, and community-centered open education through optimized OLLMs.
The research problem addressed in this study is the limited accessibility of computationally intensive AI tools in open education settings with restricted infrastructure. To tackle this, the study sets out the following objectives: (1) to evaluate the impact of pruning and RAG on system performance; (2) to analyze the educational quality of responses; and (3) to explore the feasibility of implementing dynamic fine-tuning strategies. The study is guided by three research questions: (a) How do unstructured pruning and RAG affect the response time, memory usage, and throughput of OLLMs in low-resource environments?, (b) To what extent does RAG improve the educational relevance of AI-generated content?, (c) Can LoRA-based fine-tuning mechanisms enable adaptive learning in open educational systems?
2. Conceptual Framework
The evolution of educational paradigms has arrived at a critical intersection between technological innovation and social justice in access to knowledge. The global expansion of open education and science demands innovative technological tools that are accessible, sustainable, and adaptable to diverse contexts (Chan & Lee 2023; Demeke 2023). In this scenario, artificial intelligence and OLLMs are potentially transformative. However, their high computational complexity creates a paradox: the same technologies that promise to democratize knowledge can also, if not adapted, perpetuate or even widen existing digital divides (Guo et al. 2023). This reality necessitates rethinking technology implementation strategies with an inclusive perspective.
While structured as a conceptual framework, this section also serves as a literature review by synthesizing empirical and theoretical contributions from recent studies on pruning (Guo et al. 2023; Das, Ma & Shen 2024), RAG (Lewis et al. 2021; Bura & Myakala 2024), LoRA (Chen et al. 2025), and ethical educational AI (Ferrara 2023; Camacho-Zuñiga et al. 2024). These references collectively frame the current state of knowledge and justify the exploratory-experimental approach adopted in this study.
2.1. Open Education and the Challenge of Technological Access
Open education requires much more than commitment to content availability and technological accessibility for all users. It is based on principles of universal accessibility, reusable resources, free collaboration, and justice in access to knowledge (UNESCO 2021). To meet these goals, technological infrastructures must be as inclusive as content. The dependence on high-end systems to execute educational tools is contrary to the ideals of openness, demanding adaptation strategies that allow their use in environments with limited infrastructure (Chatterjee & Bhattacharjee 2020; Ferrer et al. 2021). This technological challenge requires innovative solutions that reduce the barriers to entry for participants in open educational processes.
2.2. Computational Optimization: Pruning as an Inclusion Tool
The problem of high computational demand requires efficient solutions that do not sacrifice educational reach. Unstructured pruning of neural networks is a pragmatic solution to this challenge. Eliminating redundant parameters reduces memory, processing, and power requirements, allowing models to run on more accessible hardware (Guo et al. 2023; Navarrete & Luján-Mora 2018). This technique, initially conceived for business efficiencies, acquires new significance in the educational field: to act as a mechanism of technological justice (Chen et al. 2022). Pruning is an essential technical technique to bring educational artificial intelligence more easily to traditionally marginalized contexts.
While unstructured pruning is highly effective in reducing computational load, it is important to acknowledge alternative optimization strategies such as incremental fine-tuning. Unlike pruning, which simplifies model architecture by removing weights, incremental fine-tuning preserves model complexity while allowing continuous learning from new data, especially when paired with techniques like LoRA (Parthasarathy et al. 2024; Hossain et al. 2025). This distinction is crucial in scenarios where educational adaptability is prioritized over computational minimalism.
2.3. RAG: Guarantee of Relevance and Transparency
Optimizing models increases the need to preserve the quality and reliability of the content generated. In open education contexts, the responses must be based on verifiable, open, and contextualized sources (Adiguzel, Kaya & Cansu 2023). RAG allows search mechanisms to be integrated into open educational corpora before generating responses, improving the content’s accuracy, transparency, and pedagogical relevance (Bura & Myakala 2024; Lewis et al. 2021). In this way, using RAG strengthens the educational content and reinforces the principles of traceability and openness that characterize open education.
2.4. Open Science, Ethics, and Technological Sustainability
Finally, the incorporation of new technologies must be evaluated within an ethical framework prioritizing equity and sustainability. UNESCO Recommendations on Open Science (UNESCO 2023; Schneegans, Lewis & Straza 2021) establish the responsibility to develop open technologies that respect diversity, promote equity, and support universal access to knowledge. In this framework, the optimization of OLLMs responds to a technical imperative and an ethical need to ensure that advances in artificial intelligence fairly benefit all educational communities, without distinction of geographical location or economic capacity (Tijanić Štrok 2025). This ethical perspective reinforces the importance of designing technical solutions aligned with the fundamental principles of educational justice and social sustainability.
2.5. Incremental Fine-Tuning, Low-Range Adaptation (LoRA), and Living Educational Memory: A New Adaptive Paradigm
The concept of a living educational memory refers to an adaptive system in which user-generated educational content continuously updates and enriches a model’s knowledge base. This allows the model to evolve in real time with its learning community, ensuring contextually relevant responses grounded in traceable, open-source knowledge (Spada et al. 2023; Khatibi, Wang & Rahmani 2025).
Progress towards truly open education cannot depend solely on the availability of optimized models or accessible content. It also demands technologies that evolve dynamically with changing educational contexts. Thus, incremental fine-tuning becomes essential (Hossain et al. 2025; Parthasarathy et al. 2024). This technique allows language models to continuously adapt to new data without complete retraining (Emma 2025), significantly reducing computational costs, promoting the permanent updating of knowledge.
A key technique in this process is LoRA (Low-Rank Adaptation), which allows highly efficient fine-tuning by inserting low-rank matrices into the layers of the base model. This methodology enables substantial modifications in model behavior without altering their main parameters (Chen et al. 2025), thus decreasing memory usage and accelerating training times (Guo et al. 2023). Incorporating LoRA facilitates incremental fine-tuning sustainably (Schneegans, Lewis & Straza 2021). Access to model customization is democratized for institutions with limited infrastructure, fully aligning with open education and science principles.
In open education, incremental fine-tuning supported by LoRA improves technical efficiency and upholds a principle of cognitive justice: the ability of users to actively contribute to the construction of common knowledge. Incorporating new educational materials through a dynamic adjustment mechanism produces what can be conceptualized as a living educational memory, where each interaction strengthens, diversifies, and continuously updates the shared knowledge base (Khatibi, Wang & Rahmani 2025; Youvan 2025). Each new educational material uploaded is incorporated into the model’s response capabilities, and its provenance can be tracked, evaluated, and validated within the system, thus strengthening confidence in the traceability of the generated knowledge (Spada et al. 2023). Hence, optimizing OLLMs through incremental adjustment strategies, LoRA, and dynamic RAG is not only a technical advance but a conceptual evolution; the transformation of educational AI models into living, adaptive allies of collective knowledge construction, aligned with the ideals of equity, universal access, and sustainability of open science.
3. Methodology
This study adopted an exploratory-experimental approach grounded in recent methodological developments, integrating controlled experimentation with exploratory modeling strategies. This approach simultaneously tests optimization techniques and identifies context-driven variables that affect implementation feasibility in diverse settings (Moallemi, Elsawah & Ryan 2018; Singh 2021). Exploratory experimental design is particularly suitable for studying complex educational systems under resource constraints, where formal hypothesis testing is enhanced by iterative simulation and scenario analysis (Beauchemin & Staley 2024; Morris 2021). In this context, exploratory experimentation is not limited to the early phases of theory development but serves as a tool to generate scalable, practice-oriented evidence for educational technology design.
3.1. Selection of Models and Educational Corpus
Five open-source Open Large Language Models (OLLMs) were selected: Falcon, Bloom, GPT-NeoX, T5, and Flan-T5, recognized for their accessible licenses, adaptability to local infrastructures, and high potential for customization. These models were elaborated and refined using exclusively a corpus of open educational resources (Black et al. 2022). This strategy allowed the exploration of their ability to adapt efficiently to educational contexts, optimizing their performance without complex training processes and ensuring ethical and pedagogical alignment with open access principles. This selection sought to reflect a realistic implementation environment for universities, technical training centers, and community networks operating under open education principles (UNESCO 2023; Navarrete & Luján-Mora 2018) to establish a solid basis for evaluating the potential of OLLMs as drivers of transformation in the contemporary open education ecosystem.
3.2. Applying Optimization Strategies
The optimization of the models utilized two complementary strategies (Table 1):
Table 1
Complementary strategies.
| TECHNIQUE | EDUCATIONAL PURPOSE |
|---|---|
| Unstructured Pruning | Reduce the number of redundant parameters by 20%, allowing models to run on-premises servers at medium capacity (Das, Ma & Shen 2024). |
| RAG | Ensure that the responses generated are based on verified open educational sources, ensuring transparency and relevance (Bevara et al. 2025). |
Pruning followed the criteria of relative importance of neuronal weights (Bazanova et al. 2023; Guo et al. 2023), and RAG integrated semantic retrieval mechanisms using FAISS on uploaded open educational documents.
3.3. Simulation of Open Education Environments
Three simulated scenarios were designed to reflect the diversity of contexts in open education (Table 2) (Benítez, Van De Vijver & Padilla 2019; Ferrer et al. 2021):
Table 2
Environmental characteristics.
| SIMULATED ENVIRONMENT | REPRESENTATIVE INFRASTRUCTURE |
|---|---|
| Public University | Server with 64 GB RAM + 6 GB VRAM GPU. |
| Community Center for Digital Literacy | Mid-range laptop with 16GB RAM, no dedicated GPU. |
| Self-Organized Rural Classroom | Basic computer with 8 GB RAM and an unstable internet connection. |
Each environment ran FAQ-type educational queries, guided self-learning, and support for virtual tutors. Each optimized model was deployed and tested continuously for a period of 7 days per scenario, executing between 180 and 220 educational queries per simulated environment. These interactions covered both static prompts and dynamic retrieval tasks, recorded via system logs for performance evaluation.
3.4. Performance Evaluation
The following key variables were measured:
Response Time Reduction: Measured in percentage of improvement after optimization.
Savings in Use of Computational Resources: Percentage comparison of RAM and VRAM used.
Educational Quality of the Response: Evaluation based on adequacy, clarity, and reference to open materials.
Academic Throughput: Number of educational queries resolved per minute.
The analysis focused on determining the practical feasibility of each model, optimized for open access scenarios without compromising the quality of the educational experience. It is important to acknowledge that the study did not include direct interaction with real users (e.g., educators or students). While the simulations offer controlled insights into system performance, future studies should incorporate user feedback to evaluate educational usability, content alignment, and interaction dynamics.
Educational precision was calculated by evaluating a sample of 100 model-generated responses using a rubric based on relevance, clarity, and reference to open educational resources. Throughput was measured as the average number of user queries resolved per minute over three test intervals. All metrics were benchmarked against baseline runs prior to optimization.
4. Results
4.1. Reduction in Response Time
The unstructured pruning consistently reduced response times across all evaluated models, with improvements ranging from 10.3% to 11.5%. Figure 1 illustrates the impact of unstructured pruning on response times in four different OLLMs. Comparing original and optimized response times in this visualization highlights how computational adjustments can significantly enhance the efficiency of artificial intelligence-based (AI-based) educational platforms.
Figure 1
Reduction in response time after unstructured pruning.
As shown in Figure 1, the implementation of unstructured pruning led to a consistent reduction in response times across all models, ranging between 10.3% and 11.5%. Falcon, Bloom, GPT-NeoX, and Flan-T5 each demonstrated notable improvements, resulting in approximately 10% faster response generation. This acceleration promotes a more fluid and motivating educational experience, significant for users engaging with open learning platforms where immediacy and interaction quality are crucial.
4.2. Savings in the Use of Computational Resources
Optimizing OLLMs through unstructured pruning enabled their deployment on significantly lower-spec devices without compromising operational performance. Figure 2 depicts the reduction in computational resource consumption, specifically RAM and VRAM, after applying model optimization techniques. By contrasting original and optimized usage levels, the figure highlights how unstructured pruning can significantly improve hardware efficiency, essential for supporting open educational platforms in resource-constrained environments.
Figure 2
Reduction in resource consumption after model optimization.
As seen in Figure 2, the optimization process reduced RAM usage by approximately 19.3% and VRAM usage by 19.6%. This consistent improvement suggests that optimized OLLMs can operate effectively on mid-range or basic devices, significantly improving the technological inclusion of students and educators who may not have access to high-end computing resources. Such advancements are critical for equitable access to artificial intelligence tools in open learning ecosystems.
4.3. Improvement in Educational Quality of Responses
The integration of RAG improved response accuracy and reinforced trust in AI-generated educational content by grounding it in verifiable open sources. Figure 3 displays the improvement in educational response quality achieved through integrating RAG. The figure compares the precision rates with and without RAG, demonstrating how even slight accuracy enhancements can contribute significantly to the effectiveness of AI-assisted open learning environments.
Figure 3
Improvement in educational response quality with RAG integration.
As Figure 3 shows, incorporating RAG increased the educational precision of model-generated responses by 1.4%, rising from 94.2% to 95.6%. Although the numerical improvement may appear modest, within open educational systems handling hundreds or thousands of queries daily, this incremental gain represents a substantial enhancement of the overall learning experience. Higher response quality ensures learners receive more relevant, accurate, and trustworthy information, which is crucial for maintaining engagement and fostering more profound understanding.
4.4. Overall Impact of OLLM Optimization on Open Education Metrics
Combining unstructured pruning and RAG can generate a multidimensional improvement that makes OLLMs scalable and sustainable tools for open education systems. Figure 4 synthesizes the overall impact of the optimization strategies applied to OLLMs in four critical educational dimensions: response time, resource usage, educational response precision, and throughput. Consolidating these metrics into a single visual representation, the figure provides a comprehensive overview of how computational optimization enhances the efficiency and scalability of AI-based open learning systems.
Figure 4
Overall impact of OLLM optimization on open education metrics.
As shown in Figure 4, optimization reduced response time by 11% and approximately 19–20% in RAM and VRAM usage, significantly improving system accessibility and responsiveness. At the same time, the 1.4% increase in educational response precision ensured greater relevance and trustworthiness in AI-generated answers. Most notably, throughput (the system’s capacity to handle educational queries) increased by 33%, enabling open education platforms to simultaneously serve a much larger learner population. This convergence of gains across multiple dimensions reinforces the potential of optimized OLLMs to democratize access to high-quality education at scale.
4.5. Implementation of Optimized OLLMs in Open Education
In continuity with the proposed strategies for optimizing selected OLLMs, an experimental system integrated RAG mechanisms and incremental fine-tuning using LoRA (Black et al. 2022). This early implementation was explicitly designed to maximize the adaptability and efficiency of the models within open education environments, allowing their seamless integration with accessible and real educational materials without complex retraining or high infrastructure requirements.
The system followed the concept of “living educational memory”, in which users could upload textbooks, manuals, and open science documents that progressively enriched the knowledge base of the models. This architecture enabled extracting definitions directly from new documents, applying incremental fine-tuning on newly ingested data, and dynamically adapting the models to specific educational domains (de Fine Licht 2023). This approach is particularly relevant for its potential adoption in universities, technical training centers, and community educational networks with limited computational capacities.
Table 3 summarizes the evidence of the direct impact of this experimental strategy on three critical dimensions for the open education ecosystem.
Table 3
Preliminarily observed results.
| DIMENSION | OBSERVED IMPACT |
|---|---|
| Response Style Learning | After the initial adjustment, 100% of the responses generated consistently followed the template ‘Definition: …’. |
| Adaptability to New Content | A progressive increase in the relevance of the responses was observed, directly related to the diversity of documents that users uploaded. |
| Computational Sustainability | Each incremental fine-tuning session could be completed in less than 10 minutes using accessible GPUs (e.g., Colab T4). |
These preliminary findings solidly reinforce the hypothesis that a combination of dynamic RAG and incremental fit with LoRA can optimize the technical efficiency of OLLMs and build an open, agile, and focused knowledge ecosystem serving the real needs of educational communities (Cowling et al. 2023). This functionality anticipates an educational AI scenario where models retrieve pertinent information from open knowledge and continuously evolve by learning from the new educational materials their communities generate and use.
4.6. Comparative Summary of Model Performance
Table 4 presents a high-level comparison of the five evaluated models across the four optimization dimensions. Flan-T5 showed the best balance between educational relevance and hardware efficiency, while GPT-NeoX offered the highest throughput post-optimization. These insights can inform the deployment of specific OLLMs depending on institutional priorities, such as responsiveness, memory footprint, or output quality.
Table 4
Comparative performance of OLLMs across optimization metrics.
| MODEL | RESPONSE TIME REDUCTION (%) | RAM/VRAM SAVINGS (%) | EDUCATIONAL PRECISION (%) | THROUGHPUT GAIN (%) |
|---|---|---|---|---|
| Falcon | 11.0 | 19.0 | 95.0 | 30 |
| Bloom | 10.5 | 20.0 | 94.5 | 28 |
| GPT-NeoX | 10.3 | 19.3 | 95.6 | 33 |
| T5 | 11.2 | 19.8 | 94.8 | 29 |
| Flan-T5 | 11.5 | 20.2 | 95.4 | 31 |
Table 4 summarizes the comparative performance of the five evaluated open-source language models across four key dimensions: response time reduction, RAM/VRAM savings, educational precision, and throughput gain. Among the models, GPT-NeoX stood out for achieving the highest throughput (33%), making it especially suitable for high-volume educational platforms. Flan-T5 demonstrated the best overall balance, combining the highest reduction in response time (11.5%) with strong gains in precision and resource savings. Although Bloom and T5 offered consistent improvements across most categories, their relative performance was slightly below that of GPT-NeoX and Flan-T5. These comparative insights are valuable for guiding institutional decisions on which OLLM to deploy based on operational priorities, such as maximizing scalability, ensuring content accuracy, or minimizing hardware dependency.
5. Discussion
Reducing latency in intelligent educational systems directly affects learner engagement and task completion in virtual environments. This is supported by Figure 1, which shows an improvement of up to 11.5% in response time after applying unstructured pruning. These findings resonate with Chan and Lee (2023), who highlighted the need for real-time interaction in AI-supported education, and contrast with Luckin et al. (2022) and Bevara et al. (2025), who cautioned that speed gains must not compromise pedagogical value. This implies that future research must investigate how latency improvements impact learning retention and motivation, especially in asynchronous or self-regulated open education platforms.
Reducing hardware dependency in AI models enables broader access in under-resourced educational settings. Figure 2 supports this by showing a nearly 20% reduction in RAM and VRAM usage, making it viable to run models on mid-range or basic computers. This aligns with the inclusion frameworks proposed by Ferrer et al. (2021) and Guo et al. (2023), while challenging the assumptions of Chatterjee and Bhattacharjee (2020), who argued that technical limitations in hardware remain a fixed barrier. These results underscore the potential of model optimization as a mechanism for digital equity, suggesting that future policy and funding models in education should include technical lightness as a criterion for AI adoption.
Enhancing AI-generated outputs’ transparency and educational relevance is essential to building trust in open knowledge systems. As shown in Figure 3, integrating RAG increased content precision from 94.2% to 95.6%, ensuring more substantial alignment with verifiable educational sources. This is consistent with the open science demands outlined by Bura and Myakala (2024) and Adiguzel, Kaya and Cansu (2023), while also reinforcing Lewis et al. (2021), who emphasized the role of RAG in minimizing hallucinated outputs in generative models. For practice, this highlights the need to standardize how educational AI systems validate and cite sources. For research, it opens paths to study how students perceive and interact with AI feedback grounded in RAG pipelines.
Empowering learners and institutions to co-adapt and incrementally customize AI models marks a paradigm shift toward participatory educational technologies. Figure 4 captures this integrative gain, with a 33% increase in throughput and enhanced adaptability from LoRA-based fine-tuning mechanisms in the “living educational memory” implementation. These outcomes strongly agree with the conceptual framework of Spada et al. (2023) and Hossain et al. (2025), who view continuous learning models as foundational to cognitive justice; they expand upon Emma (2025), who emphasizes the technical side without fully addressing social co-construction. Practically, this invites the development of ethical standards for community-based fine-tuning of models, while opening a critical research agenda on how local knowledge systems can shape and be shaped by adaptive AI in education.
The ability of AI systems to evolve through user-generated knowledge inputs is redefining the role of learners as co-constructors of educational intelligence. Table 3 supports this by showing that 100% of responses adopted structured formats after fine-tuning, that content relevance increased progressively, and that each adjustment was completed in under 10 minutes using modest GPU resources. These results are consistent with Cowling et al. (2023), who argued that user involvement in training AI enhances contextual alignment, and de Fine Licht (2023), who conceptualized adaptive systems as catalysts of learning autonomy; however, they challenge Emma (2025), who treats optimization as a developer-led process rather than a shared pedagogical practice. These findings call for further exploration into the governance and curricular integration of community-trained models and suggest new research on how “living educational memory” systems may shift institutional roles in knowledge production.
5.1. Practical and Ethical Guidelines for Implementation
Based on the study’s findings, the following guidelines are proposed for key stakeholders:
Educators: Should be trained on how to customize and fine-tune OLLMs using open content from their curriculum. This ensures contextual accuracy and promotes cognitive justice.
Policymakers: Must prioritize lightweight AI deployments and require transparency mechanisms such as RAG to ensure verifiability in educational systems.
Technology developers: Are encouraged to integrate LoRA and modular training pipelines that support low-resource deployments and dynamic adaptation.These actions must be guided by principles of open science, inclusion, and ethical data use, especially in community-driven educational settings.
6. Conclusions
This study sought to determine whether Open Large Language Models (OLLMs) could operate effectively in infrastructure-limited educational contexts when optimized through unstructured pruning and Retrieval-Augmented Generation (RAG). The results demonstrated that (a) unstructured pruning reduced response time by up to 11.5%, improving system responsiveness; (b) RAM and VRAM usage decreased by around 20%, enabling execution on basic or mid-range devices; (c) RAG integration increased the educational precision of responses from 94.2% to 95.6%, enhancing content reliability; and (d) throughput grew by 33%, allowing greater scalability without performance degradation. These gains allow institutions, particularly public universities and community centers, to adopt AI-driven educational tools with limited technical infrastructure, ensuring functional access to intelligent systems. However, further testing is required to determine how these gains perform in real-world deployments, especially in diverse linguistic and disciplinary scenarios, which may affect generalizability and scalability.
The integration of incremental fine-tuning via LoRA and the dynamic use of educational documents showed that optimized OLLMs can evolve with the communities that use them. Specifically, (a) user-provided documents shaped the content of model responses; (b) fine-tuning sessions were completed in under 10 minutes using accessible GPUs, confirming technical feasibility; (c) response formats adapted predictably across sessions, demonstrating learning alignment; and (d) adaptability increased as more diverse materials were uploaded, confirming a “living educational memory” capacity. These findings suggest a new paradigm of participatory AI in education, where students and educators actively shape model behavior with local content, fostering contextualization and relevance. Yet, mechanisms to safeguard user-provided data accuracy, source validation, and ethical management remain underdeveloped and must be addressed in future studies, especially as these models scale in multilingual and multicultural environments.
The exploratory-experimental methodology applied to this study revealed technical optimization results and systemic insights into how AI can be embedded in open education. The four core results – (a) improved efficiency, (b) reduced resource usage, (c) increased educational relevance, and (d) enhanced throughput – demonstrate that OLLMs, when optimized, become realistic tools for democratizing knowledge. For practice, this means rethinking AI deployment policies in education to include lightweight, customizable, and locally governed models. For research, it means expanding beyond benchmark testing to study sociotechnical integration and pedagogical value in depth. Future work should include longitudinal analyses of learning outcomes, studies on AI–teacher interaction, and frameworks that ensure inclusive governance of models trained with community data, especially to protect underrepresented voices in the construction of open knowledge.
Ethics and Consent
This study did not involve human participants, personal data, or any sensitive information. All experiments were conducted using publicly available open educational documents. Ethical approval was not required, and no consent forms were applicable.
Acknowledgements
The authors would like to acknowledge the financial support provided by the Fondo de Apoyo a Publicaciones (FAP) of the School of Engineering and Sciences (EIC), Tecnologico de Monterrey. We would also like to acknowledge the Writing Lab of the Institute for the Future of Education, Tecnologico de Monterrey, for its valuable technical assistance.
Competing Interests
María Soledad Ramírez-Montoya served as one of the editors of this special collection but, as a co-author, did not take part in any editorial decisions relating to this paper.