Two additional baguette models released, including a 600M model said to power the Paris subway

In an October 1 thread, the author announced two more baguette models, including a 600M model they say powers the Paris subway. They claim their synthetic environment generalizes across benchmarks, with results close to Qwen 0.6B at a fraction of the training cost. They also report that epistemic calibration—recognizing what a model doesn’t know—emerges from 300M parameters onward, and say SYNTH has been downloaded half a million times.