BeijingA 35-billion-parameter AI now runs in 3GB on phones.
ModelBest says it exceeds 30 tokens a second, the minimum its co-founder calls smooth.
It predicts which expert modules the next layer needs and streams their weights from flash storage while computing.
Co-chief scientist Yao Yuan says power draw and phone-system permissions still block AI-native phones.
Yao predicts GPT-6-level intelligence will run on phones within two years, a view he says lacks consensus.
How each outlet framed it
- Zhidx leans favorable
- emphasises the ~3GB memory footprint and the 30 tokens/s on-phone speed via lookahead expert scheduling, and Yao Yuan's claim of GPT-6-class on-device intelligence within two years
Sources: Zhidx