Mianbi Intelligence's Yao Yuan: In two years, GPT-6-level AI will run on mobile phones; ordering takeout and booking flights is a bit boring now.

What happened: At a Mianbi Intelligence event on November 5, scientist Yao Yuan said the team's latest inference optimization pushed MiniCPM-V 4.7 35B-A3B generation speed on phones above 30 Token/s with roughly 3GB memory use and near-negligible accuracy Why it matters: Yao predicted GPT-6-level intelligence could run on phones with about 3GB memory within two years, a view he said lacks broad industry consensus. He said on-device multimodal capability already exceeds 2023's cloud flagship GPT-4V, and cited permission barriers and limited imagination in current AI phone use cases.

At a Mianbi Intelligence event on November 5, scientist Yao Yuan said the team's latest inference optimization pushed MiniCPM-V 4.7 35B-A3B generation speed on phones above 30 Token/s with roughly 3GB memory use and near-negligible accuracy

Yao predicted GPT-6-level intelligence could run on phones with about 3GB memory within two years, a view he said lacks broad industry consensus. He said on-device multimodal capability already exceeds 2023's cloud flagship GPT-4V, and cited permission barriers and limited imagination in current AI phone use cases.