Makes me wonder whether some Chinese labs would prefer to keep weights closed. How much is openness a preference, versus wider national pressure to compete on long-term distribution? Getting your model widely used doesn’t mean people pay you. In fact, they're more likely to pay someone else.
From @EpochAIResearch’s research on Chinese AI companies.
https://t.co/gTkkTwYsLJ released GLM 5.3 Flash’s weights. Within three weeks, its share of tokens fell from 88% to 22%. Total usage of the model there grew 17%, while its own serving volume fell 70%.
Illustrates the commercial tension:
The long-term adoption play may eventually pay dividends, but must hurt now.
Great piece, @cherylwoooo and @ansonwhho.
References
- Original post on X · 2026-10-08