While there is no clear definition of when a [[language model]] qualifies as "large", it cannot be argued that any current [[frontier model]] is *very* large, many with over 1 trillion parameters as of 2025. The dominant architecture for current LLMs is the [[GPT]]. Frontier models now incorporate vision-language models (typically with [[CNNs]]) for visual tasks and audio-language models for audio tasks. Structured data (e.g., spreadsheets) is another i/o format supported by models like [[StructGPT]]. Concerns with LLMs include hallucinations, fairness, model alignment with human values, resource efficiency, privacy and interpretability. Potential attacks on large language models include [[membership inference attack]] and [[training data extraction]]. [[frontier model]] [[open weight]] [[fine tuning]] [[vector embeddings]]