3 ms·
Theyre training the system to minimize compute,so most likely theyre dynamically downgrading quants in the first few turns hoping to find the cheapest model to
by cyanydeez 2mo ago
Theyre training the system to minimize compute,so most likely theyre dynamically downgrading quants in the first few turns hoping to find the cheapest model to run. The side effect may be excessive token gen