π-shaped Continual Learning on Qwen3.5-397B (SovereignAI claim)

Summary

Tweet-only announcement; full tech report and open-source models are “coming soon.” Schwarz (Imperial / Thomson Reuters Head of AI Research) says applying π-shaped Continual Learning to Qwen3.5-397B yields a genuine frontier model competitive with Opus 4.8 for ~$450k compute. LinkedIn copy names partners DatologyAI, Lambda, Together AI, Thomson Reuters, Imperial, and the Thomson Reuters–Imperial Frontier AI Research Lab. No method details, eval tables, or weights in the tweet (two attached charts). Unverified vendor claim until the report lands.

Keywords

continual-learning · sovereign-ai · qwen3.5 · thomson-reuters · imperial · x

Topics

continual learning, sovereign AI, post-training

Research notes

• Primary: tweet via fxtwitter (no paper/repo yet). Discord posted the tweet. Author site https://jonathan-schwarz.github.io/. Distinct from papers_local 488/506/559/612 (other CL papers). Announcement only, not a hosted corpus, so no datasets_local row. Access left blank rather than guess.