Methods, measurements and open releases.
MedErtas v0.1: domain-aware quantisation and medical drafters make open medical models about 8x faster per user at 98% to 99.6% of full-precision accuracy.
OpenRouter's 30-day data shows three structured tasks at 20.8% of classified spend, with per-token indices from 1.50x to 1.84x.
A 406M model trained on retrieved transcript spans scores within QMSum's resolution of our 1.2B system with a third of the parameters. Paper, code, models.
4-bit Gemma 4 31B build, calibrated on medical data
Perigee, the medical drafter for the 31B build: Google's Gemma 4 31B assistant drafter, fine-tuned on medical answers
4-bit Qwen3.8-27B build, calibrated on medical data
Perigee, the medical drafter for the 27B build: Red Hat's DSpark drafter, trained further on medical answers
Retrained on retrieved transcript spans.
Promoted configuration.
Protocol-exact configuration.
Collaborations, grants and press enquiries.
