A multi‑institution team released Cell2Sentence‑Scale (C2S‑Scale), a family of large language models trained on over 50 million human and mouse transcriptomes and associated annotations to represent single‑cell RNA‑seq as 'sentences.' The preprint shows C2S‑Scale can perform cross‑modal reasoning tasks — from perturbation prediction to annotation — by integrating natural language metadata with expression orderings. Developers argue this LLM framing enables richer biological queries across single‑cell datasets; they call for community benchmarking and careful validation before clinical or regulatory applications.