Nathan Lambert published a structured set of predictions on Interconnects about the trajectory of open-weight models through mid-2026 and beyond. The positions below are Lambert’s. He describes the piece as a distillation of more than ten prior articles and recordings on open models.
On the capability gap
Lambert writes that it is “surprising that the top closed models did not show a growing capability margin over open models, based on compute differences for training and research, especially in the second half of 2025 and through today.” He characterises open model labs as “technically very strong at keeping pace on well-established benchmarks,” and expects this to continue.
He also states that “closed models tend to be more robust and generally useful than similarly scoring open models” — qualities he says are not well captured by current benchmarks. This robustness advantage he sees as critical for individual-user-facing applications, such as knowledge worker assistance, where the challenge set constantly shifts.
On Chinese open-weight labs
Lambert says Chinese open-weight labs “focus slightly more on benchmark scores than comparable closed labs in the U.S.” He attributes this partly to distillation and partly to incentive structures: maintaining a narrative of keeping up with the frontier is “crucial to fundraising and adoption.” He adds that characterising Chinese labs as only benchmark-overfitters is naive — they are genuinely strong models.
His prediction is that Chinese open-weight labs will face funding difficulties “as soon as later this year,” with capability trajectory effects becoming visible three to nine months later. He stops short of predicting collapse.
On regulation and bans
Lambert writes that recurring calls to ban certain types of open models “will continue to come but are in practice impossible to implement.” His argument is that if the US bans open models above a certain compute threshold, “another sovereign entity will eventually train them and release them publicly, with the models entering the U.S. market with less oversight.”
He notes that sovereign entities and existing power structures are increasingly aware that access to highly capable AI tools cannot remain in the hands of only a small number of companies, and that this awareness will sustain long-term interest in open models regardless of regulatory pressure.
On the RL training era and closed lab advantages
Lambert argues that the current reinforcement-learning-dominated training era has created “the first clear technical area that closed labs can dominate open-weight models on capabilities.” The mechanism he describes is access to direct user feedback from deployed products like Claude Code and Codex, enabling online RL on real task distributions — something open-weight labs cannot easily replicate.
He expects open models to gain share in repetitive automation, backend automation, and AI-native applications, driven by their cost advantage. But he sees the personal assistant and knowledge worker use cases as likely to remain dominated by closed labs.
On US ground in open models
Lambert predicts that “the U.S. will slowly regain ground in adoption metrics of open models starting in early 2027,” citing Google’s Gemma 4, Nvidia’s Nemotron, and Arcee AI as examples of US-based open model efforts he considers meaningful. He acknowledges that reversing existing momentum takes time: “it takes a long time for China’s velocity to slow, then flip.”
Lambert closes by noting that the word he would apply to all of the above is “complex,” and that the long-term trajectory is “more of an economics question rather than an ability one.”