01 Evidence & Methods

Each Source Has a Different Leader

Where each chart’s numbers come from, in one card per chart. The full calculations are in the files named on each card.

The charts

  1. The most downloaded open models of the year

    What it shows
    The 15 most downloaded repositories over twelve monthly readings, October 2025 to September 2026.
    Source
    Hugging Face 30-day download counters (hub-stats, read on the 1st of each month); the ATOM Project list of open models (Apache-2.0).
    How we calculated it
    We sum the twelve 30-day readings of each first-party repository, each reading capped at five times the model’s own median month.
    Main limit
    A download is a request for a file, not a user; community conversions and fine-tunes are not added in.
    Files
    analysis/ecosystem/top_models_12m.csv · pipelines/ecosystem/top_models_12m.py
  2. Who builds the downloaded models

    What it shows
    Monthly share of downloads by the developer’s headquarters, August 2024 to October 2026.
    Source
    Hugging Face hub-stats; region labels from the ATOM Project.
    How we calculated it
    Monthly 30-day counters for the 1,676 tracked models with downloads, capped as above, grouped by the developer’s headquarters.
    Main limit
    The list is built from models that already succeeded; headquarters is not where the downloads come from.
    Files
    visualizations/ecosystem/fig1_1_data.csv · analysis/ecosystem/audit_geo_sensitivity.csv
  3. What Qwen’s downloads are made of, and the size of the models

    What it shows
    Qwen’s downloads by generation and by size band, and Qwen’s share of each size band.
    Source
    Hugging Face hub-stats; parameter counts from the ATOM Project.
    How we calculated it
    The October 2026 reading split by generation and by total parameters.
    Main limit
    Sizes are ATOM’s records; where ATOM has none, the model name.
    Files
    analysis/ecosystem/qwen_breakdown_monthly.csv · analysis/ecosystem/size_p29/qwen_lead_by_band_2026_10.csv
  4. The same developers, three measurements

    What it shows
    Downloads on Hugging Face against tokens on OpenRouter and inferences on Cloudflare Workers AI.
    Source
    Hugging Face hub-stats; OpenRouter rankings (CC BY 4.0); Cloudflare Radar, Workers AI (with Cloudflare’s permission).
    How we calculated it
    Each developer’s share in each source, for the nearest common month.
    Main limit
    Three platforms with different users and units; none is the whole market.
    Files
    analysis/ecosystem/platform_signals.csv
  5. What gets built on top, and what papers name

    What it shows
    Derivative repositories declared on Hugging Face, and research papers that name a model family.
    Source
    Hugging Face hub-stats derivatives extract (22 October 2026); OpenAlex (CC0).
    How we calculated it
    Derivatives counted by the family of their declared base model; papers by the family named in title or abstract.
    Main limit
    A declared base is the uploader’s claim; a mention is not evidence that a paper used the model.
    Files
    analysis/ecosystem/derivatives_by_family.csv · visualizations/ecosystem/fig1_2_data.csv

More data

Figures from the research that the chapter’s charts do not show.

  • Files made for local runtimes sit in separate repositories that the main download count leaves out. In ATOM’s list of such conversions, Qwen’s drew another 134.5 million uncapped downloads in the 30 days to 1 October 2026, 37% of its total once they are added, against 36% for Google’s.
  • Weight files on Hugging Face: Qwen3-0.6B, 1.5 GB; Qwen2.5-7B-Instruct, 15.2 GB; Qwen3-8B, 16.4 GB; Gemma 4 26B-A4B, 51.6 GB; DeepSeek-V4-Flash-0731, 166.9 GB; Kimi K3, 1,561 GB. Qwen and Google publish their own builds for local runtimes; none were found for the DeepSeek and Moonshot models.

Sources and notes

Show the 34 notes

Evidence levels: E1 = figure published by the source; E3 = computed by us from the source data, with no published counterpart (reproducible pipeline and tests). Every figure in this chapter is E3. Download and derivative data: cfahlgren1/hub-stats (Apache-2.0), monthly snapshots and the snapshot of 2026-10-22, revisions and checksums in data/manifests/hubstats_2026-10-22.json. Model list and region labels: ATOM Project tracked-models at commit d4b170903ec3 (Apache-2.0), data/raw/atom_tracked/d4b170903ec3/. Mentions: OpenAlex (CC0), recounted 2026-10-26, data/raw/openalex/audit_2026-10-26/. Definitions follow DEC-102 to DEC-106, each as amended by its revision 1 of 2026-10-26 (research/decisions.md). Every number in the text is listed with its source in research/editorial/sprint25/ch01-claims.csv.

  1. 1
    analysis/ecosystem/top_models_12m.csv (pipelines/ecosystem/top_models_12m.py): downloads summed over the twelve 30-day readings of 1 November 2025 to 1 October 2026 (about 360 counted days), the same population and cap as figure 1.1; developers_12m.csv for the developers' and regions' shares (Qwen 0.54; US-based developers 0.31; China-based 0.66). The window and family rankings without Qwen: top_models_windows.csv, families_no_qwen_windows.csv.
  2. 2
    visualizations/ecosystem/fig1_1_data.csv, rows 2024-08 and 2026-10, capped series (each model at five times its own median month): China-based 0.17 and 0.68; Qwen 0.10 and 0.55; US-based 0.69 and 0.30. Same values in analysis/ecosystem/audit_geo_sensitivity.csv, variant capped_5x_median (E1, baseline). List size: 1,840 rows in data/raw/atom_tracked/d4b170903ec3/models.csv (studied_ids.atom in the manifest). "About a fifth" for August 2024 covers the capped (17.2%) and uncapped (23.7%) readings (DEC-103 rev. 1). E3.
  3. 3
    qwen_breakdown_monthly.csv, 2026-10: size_band_total (under 10B 0.72, 10–40B 0.24, 40B and above 0.03, unknown 0.01), size_band_active (under 10B 0.80), architecture (dense 0.88, MoE 0.11), precision (first-party quantised 0.12; 2025-10: 0.02), specialisation (named vision-language 0.17, coding 0.04). analysis/ecosystem/qwen_size_bands.csv, basis total, Qwen's share of each band: 0.67, 0.52, 0.14 (812, 337 and 342 tracked models with parameter data). Sizes: ATOM metadata__model_parameters.csv at d4b1709 (total parameters; active where ATOM gives them, otherwise total for dense models; for Qwen3.5 and later the count excludes the vision encoder). analysis/ecosystem/qwen_top_models.csv, rank 1; qwen_concentration_monthly.csv, 2026-10: models_for_half 10, models_for_80pct 29. Image input of Qwen3.5-397B-A17B and Qwen3.8-27B: their Hugging Face model cards, read 2026-11-07. No share of all vision-language downloads is computed: the list has no modality field, and other developers' repository names do not mark it consistently. E3. Per-model rankings (section The Most Downloaded Models): pipelines/ecosystem/top_models.py, same population, cap and reading as Figure 1.1 (the 1,676 models with capped downloads above zero); analysis/ecosystem/top_models_2026_10.csv (the ten together 0.30 of tracked downloads; Qwen/Qwen2.5-0.5B-Instruct capped, raw 8,703,304), top_models_by_developer_2026_10.csv, top_models_developers_2026_10.csv (rank_of_best_model_in_all_tracked: Moonshot AI 64, DeepSeek 21). Dates: ATOM metadata__release_date_corrections.csv, otherwise createdAt in the pinned hub-stats derivatives extract of 2026-10-22, otherwise sources/model_reference_2026.csv. E3.
  4. 4
    analysis/ecosystem/qwen_breakdown_monthly.csv, dimension generation, 2026-10, share_of_qwen: Qwen3 0.45, Qwen2.5 0.24, Qwen3.5 0.15, Qwen3.6 0.07, Qwen3.8 0.06; the rest 0.02. Generation from Qwen's own repository names (pipelines/ecosystem/qwen_composition.py, generation). Growth: analysis/ecosystem/qwen_growth_by_generation.csv: Qwen's capped downloads 77,824,476 (2025-10) and 224,031,936 (2026-10); Qwen3's part of the net change 0.49. Same population, readings and cap as figure 1.1; the module reproduces 0.10 and 0.55 before computing anything. E3.
  5. 5
    Hugging Face model API (/api/models/<id>?blobs=true), each repository's config.json and model card at the pinned commit, and Qwen's, Google's, DeepSeek's and Moonshot AI's own repositories searched by name, retrieved 2026-11-08 (research/editorial/interpretation-pass28/data/hf_api/, manifest with commit hashes; scripts/fetch_hf_models.py, scripts/qwen_hardware.py, output data/qwen_hardware.csv). Parameters: ATOM's metadata, as in the rest of the chapter (Hugging Face's own count of stored parameters differs: 0.75 billion for Qwen3-0.6B, which counts the embedding table). Weight files: the summed size of the repository's weight files on the Hub. Developer's own builds: repositories under the developer's namespace whose names mark FP8, AWQ, GPTQ, GGUF or MLX; "none found" means none under that namespace on the retrieval date, not none anywhere. The 4 GB is a theoretical, weight-only figure (8.19 billion parameters x 0.5 bytes), not the memory a runtime needs. Runtimes: as named in the model cards of Qwen3-0.6B and Qwen3-8B.
  6. 6
    research/editorial/interpretation-pass28/scripts/local_formats.py, output data/local_formats_2026_10.csv: the hub-stats snapshot of 2026-10-01 (data/raw/hubstats/tracked_2026-10.csv) and ATOM's local-models.csv at d4b1709 (1,627 curated conversions, each credited to its parent), through pipelines.ecosystem.analysis.labels(). Uncapped 30-day counts, the population of the chapter's "with local conversions" sensitivity. Qwen: 228,951,136 on its tracked repositories; 26,572,840 on its own local-format repositories (AWQ, GGUF, GPTQ) and 107,956,154 on others' (GGUF 47,425,982; MLX 23,841,944, among others), 37.0% of the total. Google: 45,141,149 and 24,989,587, 35.6%. ATOM's list is a curated selection, not every conversion on the Hub, and a download of a GGUF file is a file request like any other, not evidence of local inference. E3.
  7. 7
    fig1_1_data.csv, columns share_china, share_us, share_eu: Europe-based 0.12 (2024-08) and 0.02 (2026-10). China-based share ≥ US-based share in 2024-11 (0.47 vs 0.47), 2024-12 (0.49 vs 0.46) and in every month from 2025-08; China-based share 0.65–0.69 from 2026-02 to 2026-10. "Two-thirds" and "about a fifth" are used only with this population and the capped calculation (DEC-103 rev. 1). E3.
  8. 8
    Qwen's models: 208 rows with organisation Qwen in models.csv. fig1_1_data.csv, share_qwen_of_all: 0.10 (2024-08), 0.55 (2026-10); 0.55 / 0.68 = 0.81. E3.
  9. 9
    analysis/ecosystem/audit_geo_sensitivity.csv, variants capped_without_Qwen (2026-10: China-based 0.29, US-based 0.66, 1,469 models) and capped_fixed_panel_2024-08 (2024-08: 0.17; 2026-10: 0.23). analysis/ecosystem/audit_geo_lodo.csv: 2026-10 China-based share without Google 0.77, without DeepSeek 0.66; largest move among developers other than Qwen: Google, +8.5 points. Without Qwen, 2024-08: China-based 0.09 (audit_geo_lodo.csv, row Qwen); by developer, share of the downloads without Qwen, 2024-08 to 2026-10: DeepSeek 0.02 to 0.14, Z.ai 0 to 0.07, all other China-based developers together 0.07 to 0.09 (analysis/ecosystem/audit_geo_without_qwen_by_developer.csv, Pass 26). Alternative populations, not intervals (DEC-103 rev. 1). E3.
  10. 10
    audit_geo_sensitivity.csv, 2026-10: raw 0.67; in_month_allTime_delta 0.67; with_local_conversions_raw 0.70; raw_without_month_top10_models 0.64; capped_contested_orgs_relabelled 0.68. 2024-08: capped 0.17, raw 0.24, without top 10 0.33. research/ecosystem-audit.md §2. DEC-103 rev. 1. E3.
  11. 11
    analysis/ecosystem/downloads_monthly.csv, first_party, developer mistralai: 3,052,532 (2024-12) and 60,893,655 (2025-01). Cause not verified; question prepared for Hugging Face, not sent. research/ecosystem-feasibility.md §1. DEC-103 rev. 1. E3.
  12. 12
    ATOM tracked-models README (data/raw/atom_tracked/d4b170903ec3/README.md), scope and threshold; region map metadata__organization_regions.csv (121 organisations). ATOM's eu value is a Europe-facing bucket, not a claim of EU membership. Source labels are kept apart from registry/model_openness.csv (DEC-105). Hugging Face's region:us tag denotes storage location (research/ecosystem-sources.md §2.2). DEC-102 rev. 1, DEC-103 rev. 1. E1.
  13. 13
    analysis/ecosystem/qwen_openrouter_namespace.csv, 2026-09: 0.01 (versions whose identifier begins qwen/, over the tokens of the 50 most-used versions, the "Others" bucket excluded; data/processed/openrouter/daily_rankings.csv, Chapter 03's data). Not combined with downloads (DEC-105 rev. 1). E3.
  14. 14
    pipelines/ecosystem/platform_signals.py, analysis/ecosystem/platform_signals.csv (= visualizations/ecosystem/fig1_5_data.csv): hf_share (chapter 01's capped 2026-10 reading, 1,676 models), or_open_share (September 2026 tokens of versions with an OW-P, OW-R or OW-N class in the effective registry, 64.9% of top-50 tokens; DeepSeek 0.35, Tencent 0.20, Z.ai 0.18, Qwen 0.01), cf_share (analysis/cloudflare-s11/within_task_rankings_monthly.csv, Text Generation, 2026-10; Meta 0.29). Developers are matched by namespace on each platform (DEVELOPERS in the module); a namespace is a publisher, not a model family. Workers AI offers no model from MiniMax, Tencent or Xiaomi. DEC-105 rev. 1: not combined. E3.
  15. 15
    DEC-105 rev. 1: downloads, tokens on an intermediary, mentions and bot traffic are not combined or normalised against each other. research/ecosystem-feasibility.md §2.
  16. 16
    analysis/ecosystem/audit_two_windows.csv, 2026-09, baseline: n = 35, ρ = 0.05, bootstrap 95% −0.32 to 0.37, permutation p = 0.79. Coverage: analysis/ecosystem/audit_two_windows_selection.csv (0.67; 0.32 identified with no public weights repository; 0.01 unresolved). Ranks: visualizations/ecosystem/fig1_3_data.csv. OpenRouter rankings data, CC BY 4.0 (Chapter 03). DEC-105 rev. 1. E3.
  17. 17
    analysis/ecosystem/e3_identity_review.csv (44 variants; reviewer "Claude (agent) from the captured evidence; human confirmation pending"). Confirmed only: audit_two_windows.csv, variant identity review: confirmed matches only, n = 29, ρ = 0.07, −0.31 to 0.43. Monthly baseline 2025-01 to 2026-09: ρ from −0.39 (2025-12) to 0.38 (2025-04), the only two months whose interval excludes zero (research/ecosystem-audit.md §4). Figure 1.3 status internal in design/figure-registry.json. E3.
  18. 18
    data/manifests/hubstats_2026-10-22.json, derivatives: snapshot revision a686672 (2026-10-22), 3,177,876 Hub model repositories, 902,524 declaring a base. Total: sum of repos in analysis/ecosystem/derivatives_by_family.csv and derivatives_monthly.csv, both 470,289, checked by tests/test_chapter01_s20b.py. The source snapshot lists one repository twice (neko-legends/GLM-5.3-Flash-Uncensored-EXL3); counted once, as recorded in the manifest (duplicate_ids_in_source). Earlier records give 470,290 before de-duplication. Relations are inferred by the Hub from the uploader's declaration. DEC-106 rev. 1. E3.
  19. 19
    derivatives_by_family.csv, columns repos and distinct_accounts; 229,596 / 470,289 = 0.49. Family = the base model's developer namespace, with former namespaces merged (pipelines/ecosystem/analysis.py, FAMILY_ALIASES). E3.
  20. 20
    derivatives_by_family.csv, high_volume_share: Qwen 0.34, Google 0.44, Meta 0.12; largest_account_repos for Qwen: 12,205. High-volume = an account with at least 1,000 derivative repositories across the whole extract (analysis.HIGH_VOLUME). DEC-106 rev. 1. E3.
  21. 21
    derivatives_by_year_region.csv, 2024–2026, China and US rows: repos, distinct_accounts, repos_excl_high_volume (2024 China: 80,928 − 22,353 = 58,575 from high-volume accounts). Adapters: derivatives_monthly.csv, months 2024-01 to 2024-12, base_region china, relation adapter: 73,693. Roots: analysis/ecosystem/derivatives_by_root.csv. Accounts can publish derivatives of several regions' models, so account counts are not summed across regions. 2026 runs to 22 October. E3.
  22. 22
    derivatives_monthly.csv, sum of repos_ge5_downloads: 402,104 (402,105 in research/ecosystem-audit.md, before de-duplication). E3.
  23. 23
    OpenAlex works API (CC0), title_and_abstract.search; family terms and co-term in pipelines/ecosystem/openalex.py (CO, FAMILIES). The GPT/ChatGPT rule includes ChatGPT, GPT-3.5, GPT-4, GPT-4o, GPT-5, o1 and o3; without o1 and o3, its 2025 share of all language-model works is 13.1% rather than 13.6% (research/ecosystem-audit.md §3). DEC-104 rev. 1.
  24. 24
    visualizations/ecosystem/fig1_2_data.csv, columns works_naming_any_family, all_llm_works, share_of_naming_works and share_of_all_llm_works; counts and denominators from the same day (2026-10-26). Llama, share of all language-model works: 0.02 (2023), 0.05 (2024), 0.05 (2025). Counts from 2026-10-22 differ by up to 0.3%. DEC-104 rev. 1. E3.
  25. 25
    Sum of the 14 counted families (pipelines/ecosystem/openalex_audit.py, COUNTED) over any_family, from data/raw/openalex/audit_2026-10-26/: 36,273 / 23,494 = 1.54 in 2025. Checked by tests/test_chapter01_s20b.py (test_mentions_overlap_and_ambiguous_families_stay_out). E3.
  26. 26
    analysis/ecosystem/hub_share.csv, first_party_share: 0.02 (2024-08) and 0.14 (2026-10). Selection bias: research/ecosystem-feasibility.md §1 (H1, "Sesgos"). E3.
  27. 27
    1,676: fig1_1_data.csv, column models, 2026-10 (models with a non-zero capped reading; 1,684 uncapped in audit_geo_sensitivity.csv, variant raw). 486 and 441: audit_geo_sensitivity.csv, variant capped_fixed_panel_2024-08 (note and models column, 2024-08); 390 of the 486 have downloads in 2026-10. Before the cap (variant raw), 444 models had downloads in 2024-08. The cap uses each model's median month, zero months included, so a model whose median month is zero is capped to zero; the shares do not change (Pass 26, independent review A-02). The figure of 1,840 in DEC-106 is the list size; it is not the number of models with downloads in a given month. E3.
  28. 28
    Download definition: Hugging Face, "Models download stats" (summarised in research/ecosystem-sources.md §2.1). Field availability: research/ecosystem-sources.md §2.3. Comparison: analysis/ecosystem/audit_counters.csv, column ratio_30d_over_delta: 0.89–1.02 in the 15 months other than 2025-06 (11.50), 2025-07 (0.57), 2026-05 (1.40) and 2026-06 (0.76); decreases summed from allTime_decreases (18) and allTime_decrease_sum (2026-02: −4,319,803). The cumulative counter's freeze is consistent with an import gap documented in the hub-stats README; the 2026 anomaly is unexplained and is among the questions prepared for Hugging Face (research/partner-requests-s20.md). DEC-103 rev. 1. E3.
  29. 29
    data/raw/hubstats/tracked_2026-05.csv, Qwen/Qwen3-VL-2B-Instruct, downloads 185,757,734 (snapshot of 2026-05-01). fig1_1_data.csv, 2026-05: share_china_raw 0.77, share_china 0.68. E3.
  30. 30
    derivatives_by_year_region.csv, sum of the 2023 rows across regions: 3,073. derivatives_monthly.csv, sums by path (direct 337,713; indirect 132,576) and relation. Originals that declare a base: 2,048 tracked models (manifest rows_kept 472,337 − 470,289); 362 derivatives uploaded by the base model's own developer are marked separately and included. Longest chain: manifest max_depth 16. research/ecosystem-audit.md §5. E3.
  31. 31
    research/ecosystem-feasibility.md §1 (H2, "Línea base" and "No se usa texto completo"). E3.
  32. 32
    analysis/ecosystem/openalex_precision.csv and openalex_precision_labels.csv: samples of 20 works from 2025, 16–20 with the name in the returned text; families in Figure 1.2 at 1.00 on matched works (Falcon, not drawn, 0.95); lower Wilson 95% bounds 0.81–0.84 (research/ecosystem-audit.md §3). BLOOM: 9 of 16 in model sense (0.56), excluded. Recall not measured. DEC-104 rev. 1. E3.
  33. 33
    research/ecosystem-sources.md §2.1, §1.2 and §4; DEC-102 rev. 1 (Cloudflare Radar data CC BY-NC 4.0; Interconnects/ATOM dashboard cited only) and DEC-102 rev. 2 (Cloudflare Radar's Workers AI data used in Chapter 03 with permission). DEC-101: no organisation is a research partner.
  34. 34
    research/partner-requests-s20.md (prepared, not sent). Hugging Face statistics are used as public data (editor, 2026-10-31; DEC-118). Figure statuses from design/figure-registry.json: figures 1.1, 1.1b, 1.2, 1.3, 1.4 and 1.5, provisional.