Benchmarks & research
Quyet 1.0 is Chinh Nguyen's family of five Apache-licensed decision models, from a 31B Gemma 4 model down to small encoders that run on a CPU.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
The largest model answers by reading how likely each option letter is after a fixed prompt; the quyet package then adjusts those odds so that, the maker says, 80% means right about 80% of the time. It is built for English and tuned for Vietnamese. A Gem Run example replays the maker's own timed races in both languages, where a robot picks gems on a speeding three-lane bridge.
Quyet suits sorting short messages in English or Vietnamese, such as working out what a bank customer wants and how upset they are. The small versions run on an ordinary computer; the largest needs an 80 GB graphics card. The maker has not tested it for safety or bias, so do not let it make high-stakes decisions on its own.