Every 4-6 months, a new open-weights model is released that sparks widespread debate about how closely open models are approaching the performance of the leading closed, frontier systems. The latest standout is Z.ai’s GLM 5, currently the strongest open-weights model released by a Chinese lab. Over the past year, a key development in this story is that virtually all the prominent open-source discussion models are now originating from China, whereas they used to be almost exclusively Meta’s Llama series. These discussion moments always prompt reflection for me. Despite being one of the strongest advocates for open models, I consistently find the prevailing narrative overstated: open models are not meaningfully closing the gap with the top closed models in terms of raw performance. The roughly 6-month gap remains stable. At the same time, it’s worth talking about what happens as open models continue to improve dramatically. Open models are keeping pace with the top closed models far more closely than I—and many other experts tracking the ecosystem—anticipated. On paper, the leading three U.S. AI labs—Anthropic, OpenAI, and Google—possess far greater resources for training and research. In this world, many would have anticipated a more clearly widening gap between the top open and closed models.
Interconnects AI