Schwarz claims π-shaped Continual Learning yields a frontier model for ~$450k

Jonathan SchwarzDealroom has a profile for this one. Try Dealroom → — Head of AI Research at Thomson ReutersDealroom has a profile for this one. Try Dealroom → and Associate Director of the Frontier AI Research Lab at Imperial College LondonDealroom has a profile for this one. Try Dealroom → — announced in July 2026 what he described as "a major finding for SovereignAI".

The claim: taking any open-weight model (his example: Qwen3.5-397B) and applying the team's "π-shaped Continual Learning" method is said to produce a genuine frontier model competitive with Claude Opus 4.8 for roughly $450k in compute.

Partners named: DatologyAI, Lambda, Together AI, Thomson Reuters, Imperial College London, and the Thomson Reuters–Imperial Frontier AI Research Lab.

Status: a full technical report and open-source models are said to be "coming soon". The figures and benchmarks are not yet independently verified.

Read more: LinkedIn