Blog

Notes from building Glossa: what we tried, what we measured, and what we learned.

• Can a tiny model beat Apple Intelligence?

  We fine-tuned 0.6B and 1.7B open models to explain PDF highlights and measured them blind against Apple Intelligence. The untuned models already won. Our fine-tuning taught them to cite pages, but not to be more accurate.

  November 13, 2026 · 12 minute read