It’s not memory that’s stuck on end-side AI

For the newly released mobile phones this year, manufacturers have begun to spend a lot of time saying the same thing: Our mobile phones can run large models locally without being connected to the Internet. Don’t adjust the cloud. The chip counts itself. Not only mobile phones, but also glasses and car cockpits are doing the same thing. If you talked about this two years ago, you would most likely think it was a gimmick; in 2026, it has indeed begun to become the norm. Why this year? These are three originally independent technology curves. In 2026, they will reach the same position at the same time for the first time. The first is: the model is becoming more and more "condensed". At the end of 2024, the teams of Sun Maosong and Liu Zhiyuan of Tsinghua University and the open source community OpenBMB proposed a concept called "capability density". In human terms, this is how much "smartness" a model can contain per 100 million parameters. Their conclusions are better than