Hugging Face and OpenAI

8.1  Two poles of one industry

OpenAI represents one pole: the closed-frontier model. Capability is delivered as a service over an API, the weights are retained, and what you buy is access rather than an artefact. The commercial logic is clean — the asset stays with the company, and improvements reach every customer at once.

Hugging Face represents the other: the distribution layer for everything else. Models, datasets, and demonstrations are published for anyone to download and run. It is not a lab competing on capability; it is the place the rest of the field publishes. That makes it the single best vantage point for observing what the open ecosystem actually looks like, because the platform reports its own numbers.

The two poles are not enemies so much as different answers to one question: where does the artefact live, and who can revoke your access to it? Lesson 9 turns that question into a framework.

8.2  The scale of the open ecosystem — and its concentration

Hugging Face's State of Open Source report for Spring 2026, published 17 March 2026, puts the platform at 13 million users, more than 2 million public models, and over 500,000 public datasets — roughly double the activity of a year earlier.1

Those are the numbers that get quoted. The more interesting finding is the one about the shape of demand, and the report is candid about it: supply is broad while attention is extremely concentrated. The top 200 models — on the order of 0.01% of the catalogue — account for 49.6% of all downloads.1 Roughly half the models on the platform have fewer than 200 total downloads across their entire lifetime.2

Two further structural shifts in the same report are worth more than the headline totals.

The centre of gravity moved. China has overtaken the United States in downloads on the platform, with Chinese-origin models accounting for about 41% of the total. Baidu went from no releases to more than a hundred during 2025; ByteDance and Tencent each increased releases eight- to ninefold.3

The producers changed. Industry's share of model development fell from around 70% before 2022 to 37% in 2025, while independent and unaffiliated developers rose from 17% to 39%.3 And robotics became the fastest-growing community on the platform, its datasets going from 1,145 in 2024 to 26,991 in 2025 — the single largest dataset category.3

Meanwhile more than 30% of the Fortune 500 now maintain verified accounts on the platform.2 The open ecosystem is not a parallel hobbyist economy. It is where a large share of enterprise AI work is now published and consumed.

FIG-08-01

Broad supply, concentrated demand

A long-tail distribution rendered honestly and at scale: models ranked by downloads on the x-axis, downloads on the y-axis. The first 200 entries form a towering spike shaded in the accent gold and annotated top 200 models = 49.6% of all downloads. The remaining two million stretch to the right as a nearly flat line, annotated ~half have fewer than 200 downloads, ever. Include an inset showing how far off-canvas the tail would extend at true scale. The image should feel almost absurd — that is the point.

Aspect 21:9 · 3600×1550 px · Brand palette: navy #00040c, silver-blue #a8c0d8, column gold #a87848 · Style: data visualisation, log-scale annotation permitted if labelled

8.3  Open weights are not open source

This is the distinction the lesson exists for, and it is routinely blurred in marketing.

Open weights means you can download and run the parameters. It does not mean the training data is disclosed, the training code is published, or the licence permits whatever you intend. Many widely used “open” models ship under custom licences carrying usage restrictions, acceptable-use policies, or thresholds above which separate terms apply — terms that would disqualify them under the Open Source Initiative's definition.

Why the distinction has money attached, in three practical consequences:

• Legal exposure. Deploying a restrictively licensed model commercially without reading the licence is a real liability, and “everyone calls it open source” is not a defence.
• Auditability. Without training-data disclosure you cannot fully assess bias, contamination, or provenance — which matters where you are the one accountable for the output.
• Reproducibility. Weights alone do not let you rebuild or independently verify the model. You have the artefact, not the method.

8.4  Open no longer means small

There was a period when “open model” implied “small enough to run on a workstation.” That implication is gone.

Leading open-weight models in 2026 are trillion-parameter mixture-of-experts systems requiring datacenter-scale hardware to serve. Separately there is a class of genuinely small models that run on a single GPU or a phone. What has thinned dramatically is the middle ground.4

The practical consequence for anyone reading a model card: “I can download it” and “I can run it” have become different questions. Downloading a trillion-parameter open-weight model to a laptop achieves nothing except filling the disk. Check the file sizes and the stated hardware requirements before the licence, and the licence before the benchmark scores.

Demonstration protocol

1. Walk a model card on the Hub in the §8.4 order: files, licence, data, downloads, benchmarks. Read the licence aloud — including the restrictions.
2. Show a second model card with a genuinely permissive licence and point at the difference.
3. Show the tail. Sort or search to surface a model with single-digit downloads. Figure 8.1 as a lived fact.
4. Pull a small model and run it locally. The moment a model answers on the viewer's own machine, with the network available or not, is the emotional hook for Lesson 9. Do not skip it and do not use a hosted demo for it.