US urged to embrace open-weight AI rather than isolate developers with bans

Tobi Knaup, co-founder of Mesosphere, has published an opinion piece arguing that the United States should integrate with the global open-weight AI ecosystem rather than imposing restrictions on Chinese models. Knaup draws a parallel between the current state of open-weight AI and the rise of Kubernetes, suggesting that open platforms drive superior innovation through community collaboration. The article warns that banning access to capable Chinese open-weight models, such as Z.ai's GLM-5.2 and Moonshot's Kimi K3, would isolate US developers and cede AI leadership to the rest of the world. Knaup advocates for US laboratories to release frontier-grade open-weight models and for the government to use procurement policies to support interoperable, open-source stacks.

Knaup, who co-founded Mesosphere in 2013, notes that the company was eventually disrupted by Kubernetes because the open platform became the industry’s center of gravity, allowing engineers and vendors to build upon a neutral substrate. He argues that AI is approaching a similar inflection point, where open-weight models—defined as models where trained parameters can be downloaded and modified, though not necessarily the training data—are becoming the foundation for the next AI ecosystem. While acknowledging differences, such as the lack of a neutral governance body like the CNCF for AI, he asserts that a sufficiently capable, portable substrate can attract complementary innovation far beyond what any single creator could build alone.

The gap between open and closed models is narrowing rapidly, with Z.ai releasing GLM-5.2 under an MIT license, reporting a 62.1% score on SWE-bench Pro compared to 58.6% for GPT-5.5. Moonshot has stated that its Kimi K3 model approaches closed-frontier performance on long-horizon coding tasks and has promised to publish its weights on July 27. Artificial Analysis has independently evaluated Kimi K3, scoring it alongside Opus 4.8 and GPT-5.5, supporting the claim that these models are becoming viable for the hardest coding and agentic tasks.

Despite this progress, Knaup points out that OpenAI’s and Google’s strongest models remain closed, even though NVIDIA’s Nemotron, Thinking Machines’ Inkling, OpenAI’s gpt-oss, and Google’s Gemma 4 are available under permissive licenses. He warns that if the best open-weight foundation models increasingly come from China, innovation will accumulate around them, as Hugging Face reports that Chinese models accounted for 41% of model downloads over the past year. A broad ban on American researchers and companies using Chinese open-weight models would cut the US off from an ecosystem that is already attracting many of the world’s best AI researchers and engineers.

Instead of building a wall, Knaup argues that the US should compete in the open ecosystem by releasing frontier-grade models and using procurement to create demand for portable, interoperable systems. He cites the Department of Defense’s Platform One as a precedent for using open-source tools to create demand for interoperable systems. Knaup suggests that a better approach to safety concerns is independent testing and standards for frontier models, similar to the Kubernetes conformance tests proposed by Demis Hassabis, rather than blanket bans that would sacrifice access to the entire ecosystem.