Change in AI is not only about the model that tops a leaderboard. What interests me most is the growth of an ecosystem that puts tools and weights in developers’ hands.
DeepSeek, Qwen, and Kimi offer concrete examples of this approach. DeepSeek-R1 has a public repository; Qwen3 was released in several sizes with open weights; Moonshot publishes Kimi K3’s weights and documentation under its license. The terms need reading for each individual model.
Distribution changes what is possible
When weights are available and the license permits it, private infrastructure can be considered, the model can be adapted to a specific context, and reliance on a single remote service can be reduced. These are architectural options a leaderboard does not show.
This does not mean every model is lightweight, inexpensive to run, or suitable for a company’s data. Being able to download weights and being able to operate the model in production are two different things.
A benchmark is a starting point
Results need to be read alongside the model version, configuration, available tools, and evaluation method. Measurements published by the developer are not automatically independent checks, and a good general result does not guarantee the same behavior in your own workflow.
For business software, I would compare a few real cases: response quality, errors, latency, total cost, and how easily the model can be replaced. Brand recognition and country of origin alone do not answer these questions.
Choosing AI is a decision about product architecture. The expanding open-weight ecosystem broadens that choice, but it still requires a practical evaluation.
Adapted from one of my LinkedIn posts and added to the blog on 10 September 2026. The date shown refers to publication on this website.