Open-source AI Models May Slip Down the Rankings
Benchmarks aside, the gap is growing.
Nathan Lambert has spent the past year tracking the actual performance of open source AI models. His conclusion that Chinese open models have overtaken US counterparts fed into a broader narrative of technological realignment. Continuing with our series of translated articles, here is a piece on open source versus closed models, and their changing odds in the marketplace.
Lambert has since stepped away from the scoreboard and surveyed the system. What he found was a complex landscape in which highly competitive open models are increasingly unlikely to match closed labs in every dimension.
Open models have not stalled, and they haven’t fallen behind closed models in any way that is easy to benchmark. In this respect open labs, especially those in China, continue to perform impressively. Talent is not the constraint. Nor, at least for now, is access to enough compute.
The divergence begins where benchmarks end. Closed models are consistently more robust in practice. They handle edge cases better, adapt more fluidly, and operate with a kind of versatility that is difficult to quantify but easy to feel. These qualities are becoming decisive as the use of AI shifts to agent-driven workflows. Models they do not just answer questions but complete tasks, coordinate tools, and support knowledge workers, need to work reliably.
Chinese open-source labs are operating in a competitive funding environment, which means they have leaned heavily into benchmark performance as proof of progress. Smaller or cheaper models have mimicked frontier capabilities, sustaining the narrative of “catching up with the US.” This may suffice for gathering capital and users, but it does not equate with long-term success.
Open models exist in a strange split between demand and supply. On the demand side, the case is obvious. Enterprises want flexibility. Governments want sovereignty. Developers want control. This has only grown the overall appetite for open-source systems. But who pays for their continued advancement? Unlike closed labs, which can reinvest revenue from APIs and enterprise deployments, open-model ecosystems are missing a funding loop.
Lambert suggests true divergence is a matter of resource allocation. Funding constraints may begin to bite first in China’s open-model ecosystem, and when they do, the impact will emerge over time. Investment decision differentiation will mount up over months, instead of instantly.
Closed labs are entering a new phase of advantage, driven by reinforcement learning and real-world deployment. For the first time, they are able to improve models continuously based on how they are actually used. Coding agents, enterprise copilots, and other task-oriented systems create feedback loops that are both rich and proprietary. Every interaction becomes training data. Every deployment strengthens the system. Open models, by contrast, struggle to access this kind of live, large-scale signal.
This doesn’t mean they will disappear. Quite the opposite. Open models are likely to dominate in areas where cost, control, and customization matter more than cutting-edge performance. They will power backend automation, repetitive workflows, and a wide range of AI-native applications. In these domains, “good enough” is often more than enough—and efficiency wins.
Overlaying all of this is a shifting regulatory and geopolitical backdrop. Efforts to restrict open models will likely intensify, especially as capabilities approach more sensitive thresholds. Yet such efforts may prove difficult to enforce. The barriers to training powerful models are falling, and if one jurisdiction imposes limits, others will fill the gap. Open models, by their nature, resist containment.
At the same time, the strategic importance of openness is becoming clearer. Governments and large institutions are beginning to recognize the risks of relying too heavily on a small number of closed providers. In that context, open models are not just a technical artifact. They are a governance mechanism, a way of distributing control over increasingly powerful systems.
Local agents are a sort of wildcard. Quietly, almost invisibly, a different paradigm is taking shape in which models run closer to the user, embedded in personal devices and workflows.
The future of open versus closed models is a story of interacting forces – technical, economic, and political – evolving on their own timelines.
Open models are not obviously weaker than they were before. In many ways, they are stronger. Regulation won’t stop open models, because if one country stops them, another will be permissive and release more. The U.S. may regain ground in open ecosystems starting around 2027, when momentum is likely to shift back as new entrants and platforms gain traction.
And yet, the structures that support continuous improvement are increasingly concentrated in the closed ecosystem. Unless the underlying economics of open models change, closed models are likely to pull further ahead in the ability to learn from the world, continuously, at scale, and in real time.