The Decisive Era of Open Source AI Models and Their Proliferation

Despite predictions of AI market consolidation due to rising training costs, the article reveals an unexpected proliferation of players capable of developing powerful models. Far from concentration, the market sees more and more companies investing hundreds of millions, even billions of dollars, to train models and release them as open source.

This trend is fueled by massive demand for 'tokens' and the recognition that building 'token machines' represents a significant value path. Companies like Thinking Machines, initially not perceived as open-source players, now generate hundreds of millions in revenue with their fine-tuning services and release leading open-weight models. Meanwhile, Chinese labs continue to innovate, with new entrants like Xiaomi.

The article highlights several recent open-source models, such as Inkling, Hy3, Laguna-S-2.1, DeepSeek-V4-Flash-0731, and Kimi-K3, each bringing innovations in terms of performance, licensing, or architecture. It also emphasizes the importance of licenses, such as OpenMDW or non-commercial licenses, which redefine market dynamics and geopolitical implications, particularly for American companies working with Chinese entities. We are entering an era where open models play a decisive role, challenging established economic models and market strategies.