No national open-weight model capability as a strategic hedge

The UK's frontier strategy is partnership with US labs (AISI MoU with Anthropic, Feb 2025; hyperscaler infrastructure deals). The only domestic open-model effort, BritLLM/UK-LLM (UCL, Bangor, NVIDIA; trained on Isambard-AI), is a small academic project focused on UK languages. There is no UK counterpart to the EU's OpenEuroLLM or the EUROPA consortium (selected June 2026 to build a 400B+-parameter open frontier model). Yet AISI's own trends report finds the open-closed capability gap has narrowed to 4-8 months, making maintained open weights a credible fallback the UK has chosen not to build.

If US API access were restricted or repriced, UK public services and firms have no fallback stack. Open-weight models are the one hedge that does not require winning the frontier race, but only if someone maintains, evaluates and adapts them for UK-critical functions in advance.

A National Open Models Programme under the £500m Sovereign AI Unit: a standing team that evaluates, fine-tunes, secures and hosts open-weight models on AIRR as a public-sector fallback; funding for open-model maintenance (a Sovereign Tech Fund-style instrument for model infrastructure); and formal association with European open-model consortia.

// State-led: Instrument: government programme decision within the £500m Sovereign AI Unit and AIRR.

A cheap fallback exists only if someone maintains open weights in advance; the vehicle and compute already exist, but stakes are contingent and no date forces the decision.