Post

@cxjwin The price of the model is getting lower and lower. Is this good for users and positive for industry development?

@cxjwin A flurry of open-source releases of domestically developed large-scale data models have occurred within half a month, boasting world-leading capabilities and drastically reduced costs. However, their computing power remains stuck at the previous generation. The market confuses these model breakthroughs with chip investment; the real bottleneck lies in advanced processes and manufacturing.

@cxjwin The most crucial aspect of this wave of intensive updates to domestically developed large-scale models isn't actually the number of parameters, but rather "how to squeeze performance to the limit with limited computing power." The optimization of V4-Flash's attention mechanism and the improvement in K3's computing efficiency both demonstrate how top-tier algorithms can overcome hardware bottlenecks. Solid algorithms are the foundation for developers to use high-performance, cost-effective APIs.