Evaluación comparativa de LLMs ejecutados en entorno local (ollama) para RAG académico: precisión, latencia y tasa de generación
Year: 2025
Type: article
Abstract: This article compares the performance of several language models run locally with Ollama, within a RAG system for academic assistance. Six open models were evaluated, including Ollama, Qwen, and a proprietary DeepSeek-R1 model, on tasks ranging from answering questions and solving calculations to interpreting code and generating academic text. Key metrics were accuracy, latency (time to first token), and generation speed. The findings show a trade-off: smaller models respond faster but make more mistakes; larger ones are more accurate, although they take longer. Qwen excelled with 87.5% accuracy, especially on knowledge queries, while DeepSeek-R1:8B achieved an interesting balance between speed and reasoning. Overall, the research demonstrates that combining local models, assigning each one to the stage at which it excels, can deliver efficient, cloud-independent, and accessible academic assistants.
Source: Revista vínculos
Institution: Universidad Autónoma de Guerrero
Language: Spanish
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7168116701