What About the Independents Then: Investigating Small Large Language Models in the UK

This is the authors' abstract. We don't add key points for this paper.

Using a frozen cohort of 917 model repositories and 495 publishing accounts collected on 30 October 2026, this paper investigates UK-based, from-scratch language-model builders outside university and institutional routes and examines what company-centred sovereign-AI inventories miss. Source-linked classifications identify 5 UK-based builders across 7 model entries, including 3 from-scratch builders and 1 strict independent builder. The full 495-account appendix retains institutional, overseas, derived, untrained and unresolved records. The strategic argument connects capability discovery to the race through RSI, AGI and ASI, the risk of dependent vassal-state status, and the need for a national model-building programme. The 2026 Hugging Face security incident provides a concrete example of operational dependence during an emergency. The paper addresses the National Security Strategy's baseline ambition, active defence, retained expertise, regional and distributed compute, and practical routes into model-building. The argument connects AI-assisted model development, strategic competition and deterrence, operational control of public systems, and the power, cooling, grid and regional infrastructure needed to develop domestic builders. Skills and licensing precedents inform practical routes into research. Rupert Lowe's posts of 24 October and 2 November 2026 connect sovereign AI, cyber defence and British technical expertise. The paper identifies this shared strategic concern from the leader of Restore Britain and connects it to the practical discovery and support of independent model builders.ADDITIONAL NOTESCompleted preprint with matching companion appendix, reproducible source and data. Original paper and annotations: CC BY 4.0. Original code: MIT. Third-party metadata and official statistics retain their source rights. Private logs and working notes are excluded.