ATOM Report Finds Chinese Open Models Overtook Western Peers in 2025

The ATOM Report publishes an adoption snapshot of the open language model ecosystem that covers roughly 1,500 mainline open models and reports a notable shift in market adoption: models originating in China overtook their Western counterparts in the summer of 2025 and have widened that lead since. The study mixes observable signals - Hugging Face download counts, the number of derivative models, inference market share and performance metrics - to create a composite view of who is building, using and adapting open models.

Concretely, ATOM’s dataset includes well-known families such as Alibaba’s Qwen, DeepSeek and Meta’s Llama among the population of models it tracks. Rather than presenting a single benchmark score or a launch announcement, the report looks at adoption behavior: which weights are downloaded, which models spawn derivative variants, and which attract inference traffic. That shift in focus produces a different signal than capability-only comparisons - it shows which models are actually being taken up by the community and by downstream services.

What is new in ATOM’s framing is the combination of usage and derivative metrics to detect adoption trends and a documented timing for a geographic crossover. The report finds the adoption balance flipped in mid‑2025 and that the gap has widened since, suggesting not just a transient spike but a sustained change in where community activity concentrates. By aggregating multiple public metrics rather than relying on single benchmarks, ATOM aims to surface adoption dynamics that influence implementation choices for researchers and builders.

The operational consequences of that finding are practical. Download and derivative activity matters for anyone selecting a base model to fine-tune, offering an early indicator of community-tested tooling, available adapters and compatibility libraries. For companies building inference products, models with larger derivative ecosystems typically translate to more third-party integrations, prebuilt safety adapters and deployment scripts - reducing engineering friction. Conversely, a shift in adoption can increase migration costs for teams standardized on different base families and reshape where new ecosystems of plugins, prompt libraries and deployment tooling form.

Policy and risk teams should also take note. ATOM explicitly positions its snapshot as relevant to policy advisors: observable adoption flows can guide where to focus audits, compliance checks and export-control considerations. Because the study uses public signals, it offers a transparent input to regulatory conversations about the distribution and use of open weights, although it does not and cannot account for private enterprise deployments or internal forks that are not publicly mirrored.

Context matters for interpreting these results. The report does not claim to measure every deployment or paid inference arrangement; instead, it constructs a replicable map from publicly available metrics. That makes it useful for comparing relative community attention over time, but it also leaves gaps where enterprise arrangements, closed-source derivatives or regional marketplaces operate outside the tracked channels. Readers should therefore treat ATOM as a high-resolution view of open, observable activity rather than a complete census of all model use.

Looking ahead, the most useful next steps are twofold: first, additional studies that align ATOM’s public-adoption signals with private deployment telemetry would clarify whether public attention mirrors production usage; second, market participants and policymakers should monitor whether the reported widening gap causes faster consolidation of tooling and marketplaces around the models that lead in downloads and derivatives. For researchers, entrepreneurs and advisors, ATOM provides a data - driven basis to adjust model selection and monitoring priorities - but it also underscores the need for complementary datasets to capture the full picture of where open models are actually running.