Source-linked AI summary
Economies of Open Intelligence: Tracing Power & Participation in the Model Ecosystem
Shayne Longpre, Christopher Akiki, Campbell Lund, Atharva Kulkarni, Emily Chen, Irene Solaiman, Avijit Ghosh, Yacine Jernite, Lucie-Aimée Kaffee
TL;DR
Longitudinal evidence on how power and model adoption change in the open model ecosystem has been limited. The paper analyzes Hugging Face download histories and model metadata, finding both diffusion and renewed concentration of power alongside declining transparency.
Problem
Prior research had not systematically traced how economic power or model-characteristic adoption changes over time in the open model ecosystem.
Source: §1 · p. 1Method
The authors merge Hugging Face download histories with model metadata to trace adoption and concentration across models, developers, and countries.
Source: §1 · p. 2Results
Power diffused from US incumbents toward unaffiliated and community developers, then began consolidating around Chinese developers in 2025 as model capacity grew and data transparency declined.
Source: §6 · p. 11Takeaways & Limitations
The released dataset and interactive dashboard support continued research and monitoring of concentration dynamics and changing model properties.
Source: §1 · p. 2Download metrics are imperfect proxies for real usage and are better suited to broader trends and relative popularity than precise concentration estimates.
Source: Ethics Statement · p. 12
Abstract
from arXiv · showhide
Since 2019, the Hugging Face Model Hub has been the primary global platform for sharing open weight AI models. By releasing a dataset of the complete history of weekly model downloads (June 2020-August 2025) alongside model metadata, we provide the most rigorous examination to-date of concentration dynamics and evolving characteristics in the open model economy. Our analysis spans 851,000 models, over 200 aggregated attributes per model, and 2.2B downloads. We document a fundamental rebalancing of economic power: US open-weight industry dominance by Google, Meta, and OpenAI has declined sharply in favor of unaffiliated developers, community organizations, and, as of 2025, Chinese industry, with DeepSeek and Qwen models potentially heralding a new consolidation of market power. We identify statistically significant shifts in model properties, a 17X increase in average model size, rapid growth in multimodal generation (3.4X), quantization (5X), and mixture-of-experts architectures (7X), alongside concerning declines in data transparency, with open weights models surpassing truly open source models for the first time in 2025. We expose a new layer of developer intermediaries that has emerged, focused on quantizing and adapting base models for both efficiency and artistic expression. To enable continued research and oversight, we release the complete dataset with an interactive dashboard for real-time monitoring of concentration dynamics and evolving properties in the open model economy.
1 Introduction
The introduction identifies a gap in longitudinal evidence about how power and model adoption change in the open model ecosystem. It presents a historical Hugging Face download-and-metadata analysis and reports shifting participation, model properties, and transparency.
1 Introduction
- No prior study had systematically traced changing economic power or model-characteristic adoption across the open model ecosystem over time.The introduction frames this gap against prior work on the AI supply chain, model relationships, and consolidation in closed-model markets.Source: §1 · p. 1
- The study releases Hugging Face download histories from June 2020 to August 2025, merged with metadata, to analyze adoption and concentration across models, developers, and countries.The analysis covers 851,000 models and uses download patterns as a proxy for adoption, with HHI and Gini measures of concentration.Source: §1 · p. 2
- Power shifted away from Google, Meta, and OpenAI toward unaffiliated and community developers and, by 2025, Chinese industry, while transparency declined.The introduction also reports larger models, increased multimodal generation and efficient architectures, and the emergence of developers adapting or repackaging base models.Source: §1 · p. 2
- The authors release the dataset and an interactive dashboard for continued research and monitoring of concentration and model-property changes.They say the findings provide empirical grounding for policy discussions about AI governance, economic concentration, and equitable access.Source: §1 · p. 2
2 Experimental Methodology
The methodology combines download histories with model metadata and applies filters and attribution choices to estimate adoption and concentration. It also describes sampling, annotation, and the concentration measures used.
2 Experimental Methodology
- The study merges download histories, Hub directory crawls, and manual annotations covering model derivations, architecture, modalities, methods, languages, transparency, documentation, and access restrictions.The dataset includes metadata not consistently available in automated records.Source: §2 · p. 2
- The dashboard’s leaderboard reports country, developer, and model download shares; its public data produce slightly different results from the paper’s internal deduplicated data.The display also uses symbols for model task and contributor type.Source: §2 · pp. 2–3
- The Population Sample retains 851,000 models and 97.6% of downloads; the Head Sample selects 2,875 frequently downloaded models covering 49.6% of downloads for richer annotation.The method counts one unique download per user per day and excludes models with fewer than 200 total downloads.Source: §2 · p. 3
- A one-year Rolling Window Filter counts downloads only when a model was created within a year of the download, to reduce the influence of automated loading of older models.The authors use the filter to estimate more authentic popularity and usage trends.Source: §2 · pp. 3–4
- Recursive Model Attribution assigns downloads of adapted models to their base models and developers for some analyses.The authors describe attribution to the adapted model or to its base as alternative choices with different assumptions.Source: §2 · p. 4
- Expert annotations support model-property analysis, while RANSAC estimates parameter counts from file sizes when safetensors size data are unavailable.The size-estimation method achieved validation R2 = 0.86 and raised model-size coverage from 23.6% to 98%.Source: §2 · p. 4
- HHI sums squared market shares and emphasizes leaders, whereas the Gini coefficient measures inequality across the full participant distribution.The study rescales HHI to 0–1; Gini ranges from 0 to 1.Source: §2 · p. 4
3 Concentration of Power: Open Models, Developers, & Nations
The section traces open-model power from early concentration through diffusion and renewed concentration in 2025, alongside changing model characteristics and developer participation. It also describes geographic differences in which organization types account for downloads.
3 Concentration of Power: Open Models, Developers, & Nations
- The all-time snapshot shows the United States and Western Europe dominating development, largely through industry organizations and widely downloaded text and image embedding or classification models.Examples include models based on CLIP [19], BERT [20], and YOLO [21].Source: §3 · p. 4
- Before late 2022, Google, Meta, and OpenAI commanded 40–60% of cumulative downloads, and the top 100 developers accounted for over 90%.This period was dominated by smaller encoder and embedding/classification models, with 79.3% of downloads for models disclosing training-data availability.Source: §3 · pp. 4–5
- From late 2022 to early 2024, diffusion models rose from under 1% to 20% of downloads as international, online, and unaffiliated developers gained participation.The “Rest” developer category grew from 20% to over 50% of market share during this diffusion period.Source: §3 · p. 5, Figure 2 · p. 5
- In 2025, China’s download share reached 17.1%, surpassing the United States, while DeepSeek and Qwen developers captured 14% of recent downloads and concentration began rising again.The concentration figure tracks model, developer, and national distributions.Source: Figure 3 · p. 6, §3 · p. 6
- Mean downloaded model size reached 20.8B parameters, multimodal and video generation adoption grew 3.4×, and only 39% of downloads were for models disclosing available training data.The section also reports growth in quantization and mixture-of-experts architectures, and identifies repackaging and adaptation organizations as intermediaries accounting for over 22% of recent downloads.Source: §3 · p. 6
3.1 The Rise of Communities & Unaffiliated Developers, over Industry Developers
- US, Chinese, and UK downloads skew toward industry, while other European and Asian countries and online contributors have broader participation by nonprofits, universities, and communities.Figure 4 compares the proportions of downloads allocated to developer organization types by country.Source: Figure 4 · p. 7, §3.1 · p. 7
- Industry’s share of development fell from 70% before 2022 to 37%, while unaffiliated users rose from 17% to 39% in 2025 and community developers gained share.University and nonprofit contributors remained stable in their share over the last few years.Source: §3.1 · p. 7
4 Shifts in Open Model Characteristics
This section tracks changing model characteristics across the open model ecosystem, focusing on transparency and access, model scale, and shifts toward generative multimodal architectures and efficiency methods.
4.1 The Decline of Open Source
- Training-data disclosure fell from over 79% of downloads in 2022 to 39% in 2025, while gating and restrictive licensing grew.Without training-data disclosure, a model is categorized as open weight rather than open source.Source: §4.1 · pp. 7–8
4.2 Rising Model Sizes
- 20B parameters was the mean size of downloaded models in 2025, up from 827M in 2023, while the median rose only from 326M to 406M.The large mean–median gap suggests some power users adopted multi-billion-parameter models while most deployers’ compute capacity grew more modestly; quantization and mixture-of-experts architectures are cited as enabling factors.Source: §4.2 · p. 8, Figure 5 · p. 9
- Created models’ mean size outpaced downloaded models, suggesting developers scaled faster than most deployers could adopt, with a possible lag exceeding a year.Creation-time statistics begin only in 2022 because earlier creation times were not recorded.Source: §4.2 · p. 9, Figure 5 · p. 9
4.3 The Shift to Generative Multimodal Architectures
- Downloads shifted from encoder-based discriminative models toward decoder-based generative systems, with multimodal outputs growing more varied and video generation rising 3.4×.Figure 6 shows text as 61% of weighted inputs, while outputs include text encoding (33%), text generation (22%), and image generation (19%).Source: §4.3 · p. 9, Figure 6 · p. 10
- Quantization rose 500% and parameter-efficient finetuning 40% since before 2025, while reinforcement learning and instruction tuning became more common since 2024.The authors describe a shift from techniques leveraging teacher models toward techniques leveraging reward models or synthetic data generators.Source: §4.3 · pp. 9–10
5 Related Work
Prior work examines AI power through social, economic, geopolitical, and platform lenses, while documenting that openness alone does not ensure broad access or distributed power.
Societal and Political Analyses of AI
- Critical AI studies trace concentrated compute to wealthy nations and identify technical systems as sites where social hierarchies can persist.The literature also notes contradictions in the open/closed distinction, including lower deployment costs for open models.Source: §5 · p. 10
Economic and Concentration Analyses of AI
- Prior studies report closed models account for 80% of usage and 95% of revenue, while 96% of recent AI advances originate in industry.Vipra and West argue that computational access, rather than institutional affiliation, remains the fundamental barrier.Source: §5 · p. 10
Geopolitical Analyses of AI
- Empirical studies of open-model ecosystems find that openness alone does not distribute power, because compute, data access, and expertise create hierarchies.The cited work includes analyses of China’s open-source trajectory, DeepSeek models, and open models more broadly.Source: §5 · p. 10
6 Discussion
The open model ecosystem has both diffused and reconcentrated: U.S. industry dominance fell as unaffiliated developers gained influence, followed by rising Chinese influence in 2025. Meanwhile, model capabilities and scale increased as transparency declined, and new intermediaries adapted models for community use.
Rebalancing of Power and the Cycles of Decentralization
- From 2021 to 2024, Google, Meta, and OpenAI’s download share fell from peaks above 40–60% to a marginal position by 2025, while unaffiliated developers and communities sometimes exceeded half of downloads.The shift coincided with diffusion-based generation and LoRA customization workflows that lowered participation barriers.Source: §6 · p. 11
- In 2025, DeepSeek and Qwen captured 14% of downloads, while China’s 17.1% national share surpassed the United States’ 15.7%.The authors describe this as a sign of recentralization around Chinese developers after earlier diffusion.Source: §6 · p. 11
The Rise of Intermediary Developers
- Intermediary organizations specializing in repackaging, quantizing, and adapting models accounted for over 22% of downloads in the most recent year.As average downloaded model size rose from 827M parameters in 2023 to 20B in 2025, these intermediaries focused on efficiency and deployment.Source: §6 · p. 11
Technical Transformation and Shifting Norms
- Video and multimodal generation grew 3.4×, mixture-of-experts architectures 7.2×, and quantization 5×; parameter-efficient finetuning rose by over 40%.Source: §6 · p. 11
- In 2025, the mean downloaded model reached 20B parameters, but the median was 406M, up from 326M in 2023, indicating scale growth concentrated among high-capacity deployers.The authors connect this divergence with compute constraints on typical developers.Source: §6 · p. 11
Declining Transparency
- Models disclosing any training-data information fell from 79.3% in 2022 to 39% in 2025, and open-weights models outnumbered truly open-source models for the first time.Model gating, attribution requirements, and use-condition licenses also increased; the authors note implications for reproducibility, governance, and accountability.Source: §6 · pp. 11–12
Monitoring the Open Model Ecosystem
- The dataset and dashboard provide a way for researchers and policymakers to monitor concentration, transparency, and participation longitudinally.The authors argue that static snapshots miss month-to-month changes in consolidation and participation.Source: §6 · p. 12
Ethics Statement
The statement describes limits on interpreting download counts, privacy safeguards, data and code release, labor practices, and the risks of geopolitical framing. It also records the project’s contributions and summarizes the data sources and coverage documented in the appendix.
Ethics Statement
- Download metrics are imperfect proxies for real usage, so they best support analysis of broad trends and relative popularity rather than precise concentration estimates.Source: Ethics Statement · p. 12
- The study uses public model metadata and anonymized aggregate download counts, excludes individual identifiers, and records countries only for organizations to preserve contributor anonymity.Source: Ethics Statement · p. 12
- The authors openly release the dashboard, annotation methodology, data, and code to support reproducibility, critical analysis, and further research.Source: Ethics Statement · p. 12
- Annotators were paid $25 per hour, worked without strict timelines, and could set their own schedules.The authors say they incorporated annotator expertise and perspectives into the taxonomy and acknowledge the annotators’ contributions.Source: Ethics Statement · p. 12
- The authors caution that framing AI development as geopolitical or corporate competition risks reinforcing nationalistic narratives or proprietary advantage.They encourage combining quantitative work with qualitative inquiry into social and cultural dimensions.Source: Ethics Statement · p. 12
Appendix
- Table 2 lists collected data, their sources, and whether metadata cover all models or only the Head sample.Source: Table 2 · p. 19
A Contributions •
- Project contributions included concentration analysis, data pipelines, annotation, dashboard implementation, and collection of Hugging Face download data.Source: A Contributions • · p. 19
B Data Sources
The data collection combines internal Hugging Face download logs, public model metadata, and manually or automatically inferred metadata.
B Data Sources
- The collected data combine internal Hugging Face download logs, public Hub model metadata, and manually or automatically inferred metadata.Source: B Data Sources · p. 19
B.1 Hugging Face Model Download Data
The methods describe how weekly download histories are collected, how Hub metadata are augmented, and how organization-level developer origins are assigned. A figure summarizes download shares across model attributes over time.
B.1 Hugging Face Model Download Data
- Historical download data are aggregated weekly from internal databases, and the collection is filtered to models that are publicly available at data-collection time.The public API provides only the past 30 days of download data per model; the historical series uses internal databases.Source: §B.1 · p. 19
B.2 Hugging Face Hub Model Metadata
- Metadata come from Hub APIs, organization pages, model-card data, and heuristics that infer modalities from structured task and documentation information.Figure 7 plots download shares over time for architectures, training and inference methods, and training-data disclosure and availability.Source: §B.2 · pp. 19–20, Figure 7 · p. 20
B.3 Geographical Metadata for Model Developers
- Developer country labels are assigned to organizations, not individual accounts, using evidence from organization websites and headquarters.Online or international organizations receive those labels, while unaffiliated individual contributors are labeled “Individual.”Source: §B.3 · p. 20
C Detailed Results
The detailed results figures track temporal changes in model properties, licenses, and derivations by usage. The annotation scheme supplements programmatic metadata with manual coding of model access, modalities, and language coverage.
D Annotation Taxonomy & Design
- Four trained annotators manually analyzed model information beyond programmatic metadata and model cards, consulting publications or repositories when needed.The annotators were based in the United States, Serbia, and Egypt.Source: D Annotation Taxonomy & Design · pp. 20–21
C Detailed Results
- Open Use licenses are declining as AUP licenses replace them; supervised finetuning leads derivations at 21%, followed by continued pretraining at 16%, and 7% of downloads go to models with 2+ derivation series.Quantization accounts for 4% of derivations.Source: Figure 8 · p. 21, Figure 9 · p. 21
D.1 Model Gating
- Gating annotations distinguish open access from access requiring agreement or information, including username-and-email sharing and other requirements.Access may involve accepting terms or entering information.Source: §D.1 · p. 22, Multi-label Options · p. 22
D.2 Modality
- Modality coding records input and output separately, covering single- and multimodal models as well as task outputs such as classification.Annotators use predefined labels and consult model documentation, demos, or papers when available.Source: §D.2 · p. 22
D.3 Languages
- Language annotations cover natural and coding languages, using the language of a derivative’s training process rather than its base model’s pretraining languages.When language information is unavailable on the model card, annotators may consult associated papers, dataset links, or GitHub documentation.Source: §D.3 · p. 23
D.4 Model Derivation
The annotation scheme traces which existing models supply weights to downstream models. It distinguishes inherited-weight adaptations from models that merely share architecture or code.
D.4 Model Derivation
- A model is considered derived when it inherits a pre-existing model’s weights before additional training, quantization, or other adaptation.The operational description also includes model merging among qualifying changes.Source: §D.4 · pp. 23–24
- Models using the same architecture or code without taking another model’s weights are not considered derivative.Source: §D.4 · p. 24
- Annotators record the full base-model name when known, use the exact Hugging Face model name, and mark models trained entirely from scratch as No.Source: §D.4 · p. 24
D.5 Architecture
The taxonomy records architecture, training methods, and training-data availability to characterize models and their development. For derivatives, its coding focuses on the adaptation rather than the base model’s earlier training or data.
D.5 Architecture
- Architecture coding records precise model designs, including variants and combinations such as Transformer and Mixture-of-Experts models.The taxonomy uses named categories including Transformer, ResNet, UNet, LSTM, and diffusion-based networks.Source: §D.5 · p. 24
D.6 Training Methods
- Training-method coding captures pretraining, finetuning, reinforcement learning, and other techniques, but for derivatives records only methods used to create the derivative.Examples include masked language modeling, RLHF, instruction tuning, and distillation.Source: §D.6 · p. 25
D.7 Training Data Availability
- Training-data coding distinguishes undisclosed data, partial disclosure, detailed proprietary sources, and named public datasets; derivatives are assessed on data used for the derivative.A derivative with no mention of additional data is labeled Not disclosed, even if its base model’s data was disclosed and available.Source: §D.7 · p. 26