EVALUATING LOCAL SMALL LANGUAGE MODELS FOR ARCHITECTURAL MEETING TRANSCRIPTION SUMMARIZATION AND ACTION EXTRACTION

Abstract: This paper presents a locally hosted solution, for LLM driven summarization and action extraction for projectbased meetings, that addresses the privacy, confidentiality, and cost concerns of externally hosted solutions. Developed for architectural project coordination, this system reduces the administrative burden of documenting multidisciplinary design meetings, tracking actions, and maintaining project records across distributed teams. Qwen3.6 27B was used as a judge to evaluate meeting transcript summarization and action extraction by Qwen3.5 4B and 9B SLMs in BF16 and AWQ INT4 formats. The 9B models achieved the best quality-speed tradeoff, producing accurate actions and concise, grounded summaries with lower wall-clock times.