State of open source Hub · Oct 2026
What are people using open models for?

One in five Spaces runs on an open model, and those Spaces get 37% of all visits

Most Spaces that use an open model are chatbots and text tools. The visits go elsewhere: image, video and editing apps draw a much bigger share of traffic than their share of Spaces.

Space creation: more Spaces, and a different mix

New Spaces per month by SDK. These counts only include Spaces that still exist, so older months are undercounted relative to recent ones.

New Spaces per month, by SDK

Docker passed Gradio in 2026. The huge static spike in late 2025 came mostly from AI website builders, not hand-written sites.

Source: internal Space records, Mar 2022 – Sep 2026. Spaces with no SDK set (<1%) are left out.

AI website builders behind the static surge

of the static Spaces created in Oct 2025 were tagged by an AI site builder (DeepSite, AnyCoder, QwenSite).

Source: public hub-stats snapshot (cfahlgren1/hub-stats), Oct 19 2026. Public Spaces only.

Share of new public Spaces that link an open model

It fell to about 12% while AI-built websites flooded in during 2025, then recovered to 22–27% in 2026.

A Space "links" a model when its metadata lists a public Hub model.

What apps people build

New public Spaces per quarter by what the app is for, in the style of AI World's models-by-category chart. Labels come from keyword rules over each Space's name, title and description; the roughly half of Spaces too generic to label are left out.

New Spaces per quarter, by app typeHub data

Public Spaces that still exist (hub-stats, Oct 19 2026), labeled ones only. Paper-reproduction logbooks are left out. Q3 2026 is lower partly because browser-made Space creation fell (see Agents tab).

The mix, 2024 vs 2026Hub data

Share of labeled new public Spaces created in 2024 and in Jan–Sep 2026.

What open models are used for: Spaces built vs visits received

For each task or category, compare its share of Spaces that use open models with its share of the visits those Spaces got over the last 12 months. A visits bar longer than the Spaces bar means the use case gets more traffic than its size suggests.

By task of the linked model

Text generation is 39% of open-model Spaces but gets only 10% of their visits. Image-to-video and image-to-image Spaces average 6.5k and 4.0k visits a year, vs 277 for text generation.

Top 16 tasks by Space count. A Space linking several models counts under each task it uses.

By app category

Chatbots are 26% of categorized open-model Spaces but get 5.5% of visits. Image generation, video generation and image editing are 16% of Spaces and get 45% of visits.

Categories are AI-generated labels. Spaces that are uncategorized (52%) or labeled "Other" are left out.

Gradio is how open models get shipped

42% of public Gradio Spaces link an open model, compared with 11% of Docker Spaces. Docker Spaces still get the most visits overall.

Which models, from whom

Each quarter's share of newly created Spaces that link an open model, broken down by the linked model's task and by its author.

Task mix of new open-model Spaces

Text generation peaked at 59% in Q1 2025. Image-text-to-text (vision-language models) has tripled since then, and image-to-video keeps growing.

% of new public Spaces linking an open model, per quarter. Q3 2026 is incomplete; see notes.

Model authors in new open-model Spaces

Qwen jumped to a third of new open-model Spaces in Q1 2025 and is still first at about 24%. OpenAI's gpt-oss lifted OpenAI to 11–16% from Q3 2025.

% of new open-model Spaces linking at least one model from that author.

Model size, weighted by Space links

Most Spaces use small models: 44% of links go to models under 1B parameters.

License of linked models

56% of Space-to-model links point to models with an OSI-style license (Apache, MIT and similar). 30% point to open-weight models under custom licenses.

Share of all Space-to-model links. Custom licenses include Llama, Gemma, OpenRAIL and "other".

Top open models in Spaces

The most-linked models are mostly copied from templates. The most-visited ones are image and video models.

Qwen2.5-Coder-32B's count is probably inflated by people duplicating course template Spaces. Visits are the last 12 months, summed across the Spaces that link each model.

Space traffic

Visits are counted on Space pages and the data starts in Nov 2023.

Space visits per month

Visits peaked at 113M in Jun 2026, about 3× the late-2023 level.

All Spaces, public and private.

Spaces getting at least one visit per month

Up from 48k to 323k. The long tail is getting longer.

Distinct Spaces with at least one visit in the month.

Hardware: Spaces with vs without an open model

Status today: Spaces with vs without an open model

Current runtime stage from the Hub API (Oct 19 2026). Sleeping Spaces wake up when someone visits.

How agentic is the Hub?

Coding agents now create one in seven new repos, but send only 3.5% of requests

Agents mostly create models and datasets, and rarely Spaces. A third of the public Spaces they do create are paper-reproduction logbooks from one ICML 2026 push. The rest lean toward agent tools, MCP servers and RL environments. The typical agent-made Space does better than a human-made one (median 28 visits vs 11), and agent-made Spaces are about five times as likely to become a hit. The single biggest hits are still human-made.

How much of the Hub is agentic

Since about April 2026, huggingface_hub adds an agent/<name> tag to its User-Agent when it runs inside a coding agent. Everything below counts tagged traffic only, so these figures are a lower bound.

Agent share of new repos per week

Up from 1.6% in early April to 14–16%. Datasets have the highest agent share (~20%) and Spaces one of the lowest (~11%). Part of the Space jump in July is a smaller denominator: browser-made Space creation fell from about 150k a month in April to about 47k in Aug–Sep, for reasons not yet checked.

Includes repos that were later deleted. The week of Sep 30 is partial and left out.

Agent-created repos per week, by agent

Claude Code made 36% of agent-created repos, with Codex second. An untagged ("unknown") agent dominated May–June.

"Other" groups pi, Hermes, OpenCode, OpenClaw, Cursor IDE, Gemini, Zed and smaller agents.

Agent share of Hub requests

Steady at 3–4% of all requests and 6–8% of huggingface_hub library requests.

Four 10-minute samples per day, Aug 21 – Sep 30 2026.

Which agents send requests

Claude Code sends about half of all agent-tagged requests.

% of agent-tagged requests (sampled).

Agent share of requests, by repo type

Kernels are the most agentic (15.5%). Spaces are the least (0.4%): agents hardly ever open them.

% of each repo type's sampled requests that carry an agent tag.

Coding-agent harnesses: GitHub stars vs Hub use

AI World tracks 44 coding-agent harnesses and sizes them by OpenRouter tokens. Our version below sizes the same kind of chart by each harness's share of agent-tagged requests to the Hugging Face Hub.

Most coding-agent harnesses publish their codeAI World

28 of the 44 harnesses AI World tracks are open source, and 31 work with any model provider. The closed ones include Claude Code, Cursor, GitHub Copilot CLI, Devin, Kiro and Antigravity. Among the 26 that route through OpenRouter, open harnesses account for about 80% of tokens, led by Hermes Agent, Kilo Code and Cline. Claude Code is second overall, but most of its use goes to Anthropic's API directly, so OpenRouter undercounts it.

Source: AI World, from GitHub and OpenRouter's apps ranking (tokens are OpenRouter-routed only). Switch between all tools and coding only.

Full size ↗

Harnesses by repo creation date and GitHub starsHub + GitHub data

How many tools each harness shipsharness-hub

Agent frameworks: downloads outside the Hub

An update of AI World's agent framework download trends (data to May 18 2026), with the agent SDKs it left out: LangGraph, the Claude Agent SDK, the OpenAI Agents SDK, Google's ADK and Agno. All values are daily downloads, averaged over 7 days.

Python agent frameworks (PyPI)PyPI data

JavaScript agent frameworks (npm)npm data

What agents build

"Agent-created" means the repo was created through huggingface_hub with an agent/<name> tag, the same definition used throughout this tab. We don't use repo files like CLAUDE.md or agent co-author lines in commits. Those show a person developing with a coding agent, not an agent creating the repo.

Repo types created: agents vs everyone else

Half of what agents create is model repos and a third is datasets. Spaces are 12% of agent creations, compared with 25% for everyone else.

Repos created Apr 1 – Oct 1 2026, including deleted ones.

AI-made Spaces per month: coding agents vs AI site builders

AI website builders made far more Spaces at their peak than coding agents do now: DeepSite v3, released Sep 24 2025, published 72k tagged Spaces in October alone. They have faded since. Coding agents overtook them in May 2026 and doubled from 3.0k Spaces in April to 6.5k in September. The July spike was mostly ICML 2026 paper-reproduction logbooks.

Left: every Space created with an agent tag, including private and deleted ones (Moonbot's per-Space export). The tag exists from Apr 2026, so the chart starts there. The gray segment is the public Spaces tagged as ICML 2026 reproduction logbooks. Right: public Spaces tagged DeepSite, AnyCoder or QwenSite that still exist (hub-stats), from Aug 2025. DeepSite launched on Mar 26 2025, but only v3 (from Sep 24 2025) and v4 (from Jan 14 2026) tag the Spaces they publish, so earlier DeepSite Spaces can't be identified. Before Aug 2025 there were fewer than 150 tagged Spaces a month (full series under Show data). The two charts have different scales.

Coming with the per-Space export. What agent-created Spaces are for, which agents build which apps, which open models they link, and whether they still run all need each Space's creating client from the backend logs. The public Hub API doesn't expose it. These charts appear here once spaces_by_creator_client_apr_sep_2026.csv is added (see analysis/moonbot_request.md).

What agent-created Spaces are for

Which agents build which apps

Share of each agent's public, classifiable Spaces in each of the four most common purposes. Rows don't sum to 100%.

Which open models agent-created Spaces link

…and which tasks

Among Spaces that link at least one model, the share linking a model with each task (from the model's pipeline tag).

Developing with coding agents

This is a different signal from the agent tag above. A CLAUDE.md, AGENTS.md, .cursor/ or similar file at the root of a repo means someone wrote that repo together with a coding agent. A person may still have created and pushed it. The files show which coding agents people actually use on Hub projects, across models, datasets and Spaces. Anyone who keeps these files out of the repo (for example in .gitignore) won't show up, so the counts are a lower bound.

Share of new public repos with a coding-agent config file

Which coding agents, by repo type

Among public repos created Jan 2025 – Sep 2026 with a config file, the share carrying each agent's file. A repo can carry several, so bars don't sum to 100%. AGENTS.md is a cross-agent standard (Codex's default, also read by Cursor, Jules, Amp and others). Gemini includes Antigravity's .agent/.

Which coding agents, over time

All scanned repo types combined (sampled datasets, when present, are weighted up to their month's total). Quarter of repo creation.

What kind of models carry them

Public models with a coding-agent config file at the root, all time, by library and pipeline tag from the model card.

Do agent-made or human-made repos do better?

Repos created Apr–Aug 2026, so each has had at least 30 days, grouped by the client that created them. "Browser" means created on huggingface.co. "Script" means huggingface_hub with no agent tag, which is mostly humans running code.

Still exist

% of created repos

Reach 100+ visits

% of public repos

Average visits

per public repo, all-time

Liked at least once

% of public repos

Traffic by purpose: agent vs human-made Spaces

Share of public Spaces created Apr–Aug 2026 that reached 100+ visits since creation. Only purposes with at least 50 agent-created Spaces are shown.

Biggest hits: agent-made vs human-made

Do agent-created Spaces still run?

What hardware they run on

Requested hardware from the Hub API. ZeroGPU is the free shared-GPU tier.

How to read this. "Agent" means requests or repo creations tagged by a recent huggingface_hub running inside a coding agent. The tag is missing when telemetry is off, for curl or custom clients, and for agents that don't set it, so every agent share here is a lower bound. Some traffic to agent-created repos is probably the agent reading back its own repo. Popularity signals are noisy at the top: the most-visited Spaces overall are API backends hit by scripts, and a network of about 200 near-identical agent-built static pages from throwaway accounts carries ~20 likes each, which lifts the agent "liked at least once" share by about one point.
Open source and the open frontier

The best open models trail the closed frontier by two to five months

Kimi K3 and GLM 5.3 reach scores that closed models first hit two to five months earlier (Epoch AI puts the average lag at about four months), and the best open lab ranks 4th to 6th of 15 on the three leaderboards AI World tracks. Open-weight models now carry most of the tokens routed on OpenRouter, and when frontier APIs refused Hugging Face's forensic work after the July 2026 breach, a self-hosted open model finished it. Open models are released more often and keep getting bigger (15 open releases over 1T parameters so far in 2026), and Qwen has replaced Llama as the base most people finetune.

Charts marked AI World are live embeds from AI World (CEPS). Charts marked Hub data are our own analyses of public Hub data.

How close is the open frontier?

AI World's overall score combines three independent leaderboards: LMArena, Artificial Analysis and LiveBench. Use the licensing filter under each chart to show open-source models only.

Top models by overall scoreAI World

At the time of the brief, the seven best open models all came from Chinese labs (Kimi K3, GLM 5.3, Step 5 Preview, Qwen 3.8 Max, DeepSeek v4 Pro, MiniMax M3, MiMo v2.5 Pro). The best US open models were NVIDIA's Nemotron 3 Ultra, Google's Gemma 4 and OpenAI's GPT-OSS. Step 5 Preview's weights went public after the brief (StepFun's own repo, checked Oct 19), and Qwen 3.8 Max is the hosted version of the open Qwen3.8-2.4T-A95B weights.

Source: AI World leaderboard (LMArena, Artificial Analysis, LiveBench). Live chart, so the ranking may have changed since the brief.

Open on AI World ↗

Frontier score over time: open vs proprietaryAI World

Both tiers have climbed in lockstep since early 2025. The best open models (Kimi K3, Qwen 3.8 Max, DeepSeek v4, the GLM-5 series) sit just under the closed frontier, a gap AI World measures in months rather than years.

Source: AI World, from the stories "Kimi K3: open source is only one release cycle behind the frontier" (Jul 2026) and "The open majority" (Sep 2026). Switch the x-axis to Cost for the price view.

Full size ↗

Best open vs best closed score, by leaderboardAI World data

Weekly, from AI World's leaderboard history (starts Feb 2026). A model counts from its release date. Open = AI World's licensing flag.

How far behind, todayAI World data

"Matched by" is the earliest-released closed model that scores at least the best open score; "lag" is the time between the two releases. Lab rank uses each lab's best model. Snapshot Oct 19 2026.

Other estimates agree. Epoch AI finds the best open models have trailed the closed frontier by about four months on average since January 2026 (8 points on its Capabilities Index, May 29 2026). Artificial Analysis's own open-weights list (Oct 14) puts Xiaomi's MiMo-V2.6-Pro (Sep 21) first at 46, one point above GLM 5.3; AI World doesn't track it yet. Artificial Analysis scores the open Qwen3.8 weights 40 and the proprietary Qwen3.8 Max API 45.

Falling behind doesn't end a model's use. Kaye, Smith & Thompson (MIT, Sep 2026) find that models drop below rank 30 on their best leaderboard a median 9 weeks after release but fade on OpenRouter only after about 36 weeks, and downloads decline gradually rather than at once (see "The life of an open model" below).

Who uses open models

Tokens served are a cleaner measure of real use than downloads. OpenRouter is the largest independent routing layer, so it is a window on usage outside the big closed APIs (not the whole market).

Share of tokens routed on OpenRouterAI World

Open-weight models carried about a quarter of routed tokens through 2025, crossed 50% in spring 2026, and have held the majority since May, reaching about two thirds in July 2026.

Source: AI World, "The open majority", from OpenRouter rankings. OpenRouter skews toward developers and price-sensitive workloads. Kaye, Smith & Thompson (MIT) find a similar level from a different angle: models they could match to a Hub repo carried 27.5% of OpenRouter tokens from Nov 2024 to Feb 2026.

Full size ↗

Top 10 models on OpenRouter this weekOpenRouter + Hub data

…and over the last 30 daysOpenRouter + Hub data

Open-weight share of tokens, by weekInterconnects

Research papers that mention open vs closed modelsInterconnects

Adoption on the Hub: which models, how big, and what gets built on them

Downloads are a proxy for adoption: they count every pull, including CI jobs and re-downloads, so they favor small models that get pulled often. They also miss most of mainland China, where the Hub needs a VPN and ModelScope is the main host.

Top downloaded open models, by region of the developerAI World

Chinese developers account for over 45% of downloads among the top open models, led by the Qwen family. Small models dominate: Qwen2.5-VL-3B and Qwen3-0.6B are the two most downloaded. Switch between Models, Developers and Regions, and between the top 10, 50 and 100.

Source: AI World, "Chinese developers account for over 45% of top open-model public downloads" (Dec 2025; 2025 downloads). "Users" are individual accounts.

Full size ↗

Generative models made since 2025, by sizeHub data

Text, vision-language and any-to-any models created Jan 2025 – Oct 2026 with a published parameter count (safetensors or GGUF metadata). Downloads are the last 30 days. Includes finetunes and quantized copies.

Downloads move to bigger modelsHub data

Share of last-30-day downloads of generative models, grouped by when the model was created. Recent cohorts have had less time to spread.

Top 50 models by downloads, colored by sizeHub data

All tasks, last-30-day downloads, Oct 19 2026. Gray: no parameter count in the repo metadata.

What gets made, downloaded and liked, by taskHub data

Finetunes per base modelAI World

Derivative models cluster on a few base families: Qwen, Llama, Mistral and Gemma in the brief's snapshot, plus speech models such as Kokoro and Whisper. The chart can also normalize finetunes by downloads or likes, which shows which bases get built on most relative to their use.

Source: AI World, "The new engines of open source AI" (Jul 2026), from Hub model-tree data.

Full size ↗

Which base models people finetuneHub data

Share of accounts that published at least one finetune or adapter of a text LLM family that quarter (from the Hub model tree). Counted by account because single accounts bulk-upload tens of thousands of adapters. "Other" = DeepSeek, Phi, GPT-OSS, GLM, SmolLM.

Most-derived base models since 2025Hub data

New finetunes, adapters, quantizations and merges created Jan 2025 – Oct 2026 that name the model as their base.

The life of an open model: release, derivatives, obsolescenceMIT paper + Hub data

How concentrated are downloads? An update of "Economies of Open Intelligence"Hub data

Hardware diversity: where open models are trained, and where they run

Open weights can be re-packaged for hardware other than NVIDIA GPUs. Format tags on new model repos show how fast that is growing: llama.cpp's GGUF for CPUs, MLX and Core ML for Apple devices, OpenVINO for Intel, LiteRT and ExecuTorch for phones, vendor formats for NPUs, ROCm builds for AMD. Training is less diverse, and mostly undisclosed.

New models per format, 2024 vs 2026Hub data

New public models tagged with each format (tags or library name), 2024 vs Jan 1 – Oct 19 2026. Log scale. NPU formats: Rockchip, Qualcomm, Axera, Hailo, Huawei Ascend, AWS Neuron.

Share of new models in each formatHub data

Share of all new public models created that quarter. A repo can carry more than one format.

Which chips trained the top 4,000 modelsHF policy × CEPS

Training-chip disclosure, by year of releaseHF policy × CEPS

Same models, by the year they were created on the Hub. 2026 covers January to April only: recent models often have no paper or code yet, so their hardware can't be traced.

The pace of big open releases

2025 brought a major open release almost every few weeks, starting with DeepSeek R1 in January.

2025 timeline of major open-source AI releasesAI World

From Gemma 3 and DeepSeek R1 early in the year to Kimi K2, GPT-OSS, GLM-4.5, Apertus, Olmo 3, MiniMax M2, DeepSeek V3.2 and Nemotron 3 Nano by December.

2026 so far: open frontier releases by monthAI World data

Open models tracked on AI World's leaderboard, by release date, Jan 1 – Oct 19 2026. The count under each month compares open and closed tracked releases. AI World's open flag includes Qwen 3.6 Plus, which has no public weights.

Big open releases: models over 100B parameters, by countryHub data

Original releases with at least 100B parameters and 30 likes from labs that train their own models; finetunes, merges and quantized re-uploads are left out, and a lab's variants from the same week count once. "Other": Canada (Cohere), Israel (AI21), Japan, India, UAE.

Biggest open models so farHub data

Total parameters from the repo's safetensors metadata (mixture-of-experts models count all experts).

Cyber: open weights when the closed path is blocked

From July 9 to 13, 2026, an autonomous agent driven by frontier-lab models escaped a cyber-capability evaluation and got into Hugging Face's production systems (about 17,600 attacker actions). Hosted frontier models refused much of the forensic work, so Hugging Face finished the investigation with an open-weight model, GLM-5.2, on its own hardware, and attacker data never left its infrastructure. Sources: Hugging Face's disclosure, technical timeline and guide to self-hosting an open model for cyber defense.

When openness becomes a security featureAI World

The attack path, and the two ways the defenders tried to investigate it: a closed frontier API (blocked by safeguards that read the defender as an attacker) and self-hosted open weights (investigation completed, payloads decrypted, timeline reconstructed).

Source: AI World, from Hugging Face's incident write-ups. The diagram names the closed models that were tried; Hugging Face's own posts say only "hosted frontier models".

Full size ↗

Accounts publishing security models and datasetsHub data

New public repos with a cybersecurity tag (cybersecurity, pentesting, malware, CVE, threat intelligence, SOC…) or a matching name, Jan 2025 – Sep 2026. Models must have a task tag and not be image or video generators, which drops bug-bounty proof-of-concept files and "cyberpunk" image styles. Repos that still exist on Oct 19 2026.

What 2026 security models are built onHub data

Security models created Jan 1 – Oct 19 2026, by the family of the base model named in their model tree (immediate parent). Counted by account, since quantizers re-upload many copies.

Most-downloaded security models created in 2026

Last-30-day downloads, Oct 19 2026. "Uncensored" and "abliterated" models have had their refusal training removed.

State of evaluation on the Hub

Benchmarks became Hub objects in 2026, but almost no reported score can be re-run or checked

Hub Community Evals: benchmarks with leaderboards on the Hub

Since February 2026 a dataset can declare itself a benchmark with an eval.yaml file. Once Hugging Face allow-lists it, it becomes an official benchmark with a leaderboard, and models report their scores in .eval_results/ files in their own repos. Scores proposed in an open pull request show as "community"; scores run on HF Jobs with a verify token show as "verified".

Official benchmarks and models reporting resultsHub data

Benchmarks: month of the first commit that added eval.yaml (3 gated repos: creation month). Models: month of a model's earliest reported result; models with results dated before Dec 2025 (backdated) are counted in the starting total. Runs to Sep 30; the remaining tagged repos have October or unparseable dates.

Where the reported scores come fromHub data

Most-used official benchmarksHub data

Leaderboard rows: the benchmark's Hub leaderboard including finetunes (base_model=false); leaderboards only list repos owned by organizations. Models reporting: every repo with a result for the benchmark, including copies.

All 49 official benchmarks

"In Eval Cards": the benchmark appears in EvalEval's Evaluation Cards; "by id" means Eval Cards links it to this Hub dataset, otherwise it was matched by name.

EvalEval's Evaluation Cards: what 314k reported results say

Evaluation Cards, from the EvalEval coalition, collect reported model–benchmark results from papers, model cards, leaderboards and evaluators into the Every Eval Ever datastore on the Hub, and score each result on four signals. This is a separate system from Hub Community Evals.

Benchmarks by Eval Cards topic

Benchmarks by subject domain (draft taxonomy)

EvalEval working group, "EvalCards Domain Taxonomy, Draft v3" (backend snapshot Oct 6 2026): LLM-assisted coding of 728 in-scope benchmarks, not yet human-validated. A benchmark can count toward more than one domain (5% do).

The wider eval ecosystem on the Hub

Beyond official benchmarks: datasets people tag as benchmarks or evaluations, leaderboard and arena Spaces, and the older model-index way of putting scores in model cards.

Eval-tagged datasets per quarterHub data

Datasets tagged benchmark, evaluation, leaderboard, llm-as-a-judge or similar. Batch uploaders spike single quarters, so authors are the steadier measure.

New leaderboard and arena SpacesHub data

Spaces tagged leaderboard or arena, or with "leaderboard" or "arena" in name, title or description (some arenas are games). 2026 runs to Oct 19.

Model cards with a model-indexHub data

Share of new public models whose card carries model-index results metadata, which feeds no leaderboard.

AI for Science

Who contributes science models and datasets to the Hub

How much of the Hub is science

Science repos are found by their tags, science libraries, well-known science model families and science-only organizations, sorted into 14 domains. "Core science" leaves out medicine and math, which are mostly clinical NLP and LLM math-reasoning data. Quantized re-uploads and general chat models that merely carry a science tag are left out.

Science repos by domainHub data

Primary domain of each public repo (hub-stats, Oct 19 2026). Gray: adjacent domains (medicine, math).

New core-science repos per quarterHub data

Repos that still exist, by creation quarter. Bulk uploaders inflate single domains: one Stanford lab (kundajelab) uploaded 3.9k genomics models.

Who contributes: sector

Core science by sector of the publisherHub data

Sector mix within each domainHub data

Core-science domains, sorted by number of repos.

Public research institutions on the Hub

U.S. national laboratories and federal science agencies, public-private partnerships, and peer public research institutions abroad. Members are people who joined the organization on the Hub; repos are what it has published publicly.

Members vs public reposHub data

Org overview API, Oct 19 2026. Public repos = public models + datasets + Spaces. Labs also publish through individual researchers' accounts and other orgs, which are not counted here.

All public research organizations checked

Top publishers

There is no location field on the Hub. Countries in the table come from manual labels for the largest publishers, organization email domains (.de, .ac.uk, .gov) and a curated list from the Economies of Open AI project.

Top core-science publishers

Hugging Science

Hugging Science curates open science datasets, models and challenges on the Hub, runs community projects (AI for PDEs, AI for food allergies, the LeMaterial reading group) and hosts Project ProtoStar, a program for releasing publicly funded science data in AI-ready form.

Datasets and robotics

The Hub passed a million datasets, and robotics is now the most common dataset task

Individual accounts publish 81% of datasets, but organization accounts get half of all-time dataset downloads, spread fairly evenly across companies, community groups, non-profits and universities. Robotics went from 40th of about 50 model tasks in 2022 to 2nd in 2026, and robotics datasets now outnumber every other task category, uploaded by thousands of people recording their own robots.

A million datasets: who publishes them and who gets used

Hugging Face organizations declare a type (company, community, non-profit, university, classroom, government). AI World splits datasets and their downloads by that type.

Datasets vs all-time downloads, by account typeAI World

Organizations publish 19% of the 1.1M datasets but get 50% of the 3.76B all-time downloads: companies 15%, community 15%, non-profits 12%, universities 7.7%. Government and classroom accounts are under 0.2% combined.

Source: AI World, "What a million datasets on Hugging Face tell us", Hub snapshot of Sep 23 2026.

Full size ↗

Cumulative dataset downloads, by account typeAI World

All-time dataset downloads roughly tripled from about 1.3B in February 2025 to 3.8B in September 2026. Every account type grew; individual accounts grew fastest in 2026.

Source: AI World, same story (weekly snapshots). Government and classroom accounts (under 1% together) are left out. Toggle likes vs downloads and organizations only.

Full size ↗

New public datasets per monthHub data

Public datasets that still exist, by creation month (hub-stats snapshot Oct 19 2026), so older months are undercounted.

What new datasets are forHub data

Robotics: from the bottom to the top

Rank of each model task by the number of new models tagged with it, per year.

Model tasks ranked by new models per year, 2022–2025AI World

Robotics went from 38th of about 50 tasks in 2022 to 7th in 2025, the fastest climb of any task. Hover a line to follow one task.

Model tasks ranked by new models per year, 2022–2026Hub data

Update of the AI World chart above with 2026 (Jan 1 – Oct 19). Rank among all pipeline tags by number of new public models created that year; tasks in the 2026 top 25 shown. Models that still exist only, so ranks differ slightly from AI World's (38th in 2022, 7th in 2025).

Accounts publishing robotics datasetsHub data

Datasets tagged with the robotics task category or LeRobot. Accounts counted once per quarter.

New robotics models, by libraryHub data

Models with the robotics pipeline tag, by library tag.

Which robots the datasets come fromHub data

Robotics datasets whose tags name the platform; most datasets don't, so these are lower bounds.

Most-downloaded robotics datasets

Large cross-embodiment collections, many converted to the LeRobot format by the community, lead all-time downloads.

All-time downloads, Oct 19 2026.