Towards sustainable AI knowledge-base assistants in computer science education: on-premise deployment and optimization with open educational resources

Synthesis: Shen et al. (2026) demonstrate the feasibility of running AI knowledge-base assistants entirely on consumer-grade hardware (NVIDIA RTX 3060, 12 GB VRAM) using open educational resources (OER) for computer science education — an on-premise alternative to cloud-dependent AI tutors. Combining structured extraction from 82 OER documents, RAG, and NF4 4-bit quantization-aware fine-tuning of Qwen-7B and DeepSeek-MoE, they report strong accuracy (DeepSeek-MoE 78.6% overall, 82.3% multi-hop) at very low energy cost (1.8 mWh/query) while preserving educational utility within ~2 points of full-precision accuracy. The system is framed as a knowledge-base assistant, not a validated tutor.

Key Findings

1. On-premise deployment is feasible on consumer GPUs. AI inference runs entirely on institution-owned hardware, eliminating cloud latency, cost, and data-Governance risk — key for regions with unreliable internet or limited infrastructure.
2. Retrieval augmentation is essential. The Local LLM (no retrieval) scored only 52.3% accuracy — below even the TF-IDF baseline (55.4%); RAG without fine-tuning reached 66.6%.
3. Fine-tuning adds value beyond retrieval. Qwen-7B (quantization-aware fine-tuned) reached 69.8% (+3.2 pp, p=0.031); DeepSeek-MoE reached 78.6% (+12.0 pp, p<0.001), including 82.3% on multi-hop reasoning.
4. Quantization-aware fine-tuning preserves utility. It narrowed the FP16-to-4-bit accuracy gap to 1.7 pp (Qwen-7B) and 1.2 pp (DeepSeek-MoE), while cutting VRAM by ~38% and energy to 1.8 mWh/query (43.8% below baseline).
5. Hallucination is controllable. Quantization raises hallucination rates, but fine-tuning substantially mitigates them (e.g., DeepSeek-MoE 10.4%→8.1%), measured via a two-stage NLI procedure against retrieved OER chunks.

Implications

This work connects open-source AI, Privacy (on-premise data control), and CS Education to sustainable, resource-constrained deployment. It grounds AI Ed Evaluation in a multi-dimensional protocol (accuracy, efficiency, RAG faithfulness, energy) rather than single accuracy numbers — a model for how institutions with limited budgets can evaluate and adopt local AI tutors. It also aligns with Benchmark-driven comparisons and retrieval-augmented generation design, while the explicit "knowledge-base assistant, not validated tutor" framing cautions against overclaiming pedagogical effectiveness.

Connected Concepts

• CS Education
• Open Source
• Privacy
• LLM
• RAG
• Benchmark
• AI Ed Evaluation
• Generative AI
• Edtech Platform
• Intelligent Tutoring
• Sustainability

Connected Articles

• Vismatic Secure Sandbox CS Education — VS-MATIC: secure sandboxed environments for computing education
• Lata Ferpa Compliant Local LLM Autograder — FERPA-compliant local LLM grading
• Oatutor Open Source Adaptive Tutor 2023 — Open-source adaptive tutor
• Simulating Students Java Programming Errors Llms — LLMs as simulated learners in CS
• Reshaping CS Education GenAI — Reshaping CS education with GenAI
• Caruana Pre University AI Education Slr 2026 — SLR of pre-university AI education (Caruana et al. 2026)

Citation

Shen, X., Feng, L., Hua, S., Liu, D., Xie, Z., & Liu, B. (2026). Towards sustainable AI knowledge-base assistants in computer science education: on-premise deployment and optimization with open educational resources. Frontiers in Psychology, 17, 1843444. https://doi.org/10.3389/fpsyg.2026.1843444