Hugging Face Was a Chatbot for Teenagers When It Ported BERT to PyTorch

Transformers lets developers run models other people trained

Training a large language model takes weeks of computing time and data most teams do not have. Using one that someone else trained takes a few lines of code, if the code exists. Transformers, a Python library, is that code for thousands of models. A developer names a model, such as bert-base-uncased, and the library downloads the trained weights, loads them into the right network and returns something that can classify text, answer a question or write a reply. The models live on the Hugging Face Hub, a website where companies and researchers publish them. The library is free and Apache-licensed. Hugging Face sells hosting, compute and enterprise accounts around it.

Hugging Face started as a chatbot for teenagers

Clément Delangue and Julien Chaumond founded Hugging Face in New York in 2016, named after the 🤗 emoji. Thomas Wolf, now called the third co-founder, said in 2021 that Chaumond “invited me to join and build their science team”. The company went through the first Botcamp, a program for chatbot startups run by the investor Betaworks, which says it focused on “NLP and conversational interfaces”. In March 2017, TechCrunch reviewed the iOS app as “a new chatbot app for bored teenagers” and reported $1.2 million raised “from Betaworks, SV Angel and NBA star Kevin Durant and others”. Delangue had told the reporter during the private beta: “We’re building an AI so that you’re having fun talking with it.”

In May 2018, TechCrunch reported a “$4 million seed round led by Ronny Conway from a_capital”. The app then handled “1 million messages per day”. In December 2018, the homepage still read: “The social AI who learns to chit-chat, talks sassy and trades selfies with you.”

The team already published code. Its coreference library, huggingface/neuralcoref, and its emotion model, huggingface/torchMoji, came out in 2017, and in June 2018 it released a PyTorch version of OpenAI’s first GPT model. Those repositories have 2,890, 922 and 1,525 stars today.

Three people ported BERT in a week

Google published the BERT paper in October 2018, with code in TensorFlow. Delangue later recalled Wolf’s view of it: “It’s amazing, but it sucks because it’s in TensorFlow.” Wolf created pytorch-pretrained-BERT on 29 October, two days before Google pushed its code on 31 October. Over the next week, Wolf, Victor Sanh and Tim Rault made 123 commits: 62 by Wolf, 44 by Sanh, 15 by Rault and 2 from a server account. On 5 November, Wolf posted: “Here is an op-for-op @PyTorch re-implementation of @GoogleAI’s BERT model by @SanhEstPasMoi, @TimRault and I.” The post shows 626 likes today. The first release followed on 17 November.

Delangue tells the story as a weekend. On the ACQ2 podcast in 2024 he said Wolf told the founders on a Friday that he would spend the weekend porting BERT, and “On Monday he released a PyTorch version of BERT, tweeted about it, and I think his tweet got maybe a thousand likes.” The commits show three people working from Monday 29 October to Monday 5 November.

Hacker News barely noticed the launch

A Hacker News post on 7 November drew 2 points. GitHub dates 369 surviving stars to 5 November, the day of Wolf’s post, and 1,114 to the first ten days.

GitHub star history for huggingface/transformers, 3 to 14 November 2018, collected on 12 November 2026. Removed stars are not counted.

These counts are floors. An archived copy of the repository page showed 1,316 stars on the evening of 14 November, when at most 1,114 of the stars dated by then still exist. Within three weeks, the library had passed the 1,525 stars the GPT port has today. Launch day is the repository’s fifth-best day. The best is 17 July 2019, with 480, the day after the library was renamed.

Other people’s models turned it into a library

Release 0.5.0, on 11 February 2019, added OpenAI’s GPT and Transformer-XL, and 0.6.0 added GPT-2 a week later. In July 2019, version 1.0 added XLNet and XLM and renamed the project pytorch-transformers. In September, version 2.0 added TensorFlow 2.0 support and the name became transformers. Delangue said that people “told us that they wanted to add their model”, and that this was what took it “from a single model repository”.

The homepage changed before the funding did

Delangue recalled: “I think six months later we told our investors, look, this is adoption. This is the usage that we’re getting on this new platform. We think we need to pivot from one to another.” He added: “For us it was three years in with maybe $6 million raised.”

The homepage shows the change happening slowly. In February 2019, it listed “PyTorch BERT 3k+ stars on GitHub” among its science projects, under a pitch for the app. On 9 August 2019, it still asked: “Are you ready to talk with the most hilarious robot on the block?” By 3 September, the headline was “On a mission to solve NLP, one commit at a time.” The page title still said “Fun chat with your own Artificial Intelligence” until at least 4 September, and an app called Talking Dog was still promoted below the new headline. We found no source that says when the chatbot app was retired.

Lux led the Series A and cited the stars

On 17 December 2019, Hugging Face announced a $15 million Series A led by Lux Capital. Brandon Reeves of Lux wrote: “They’ve won the open source community with over 1 million installs, ~19,000 Github stars, and many use cases in production. Hugging Face is one of the fastest-growing open-source projects we have ever seen”. His post includes a chart captioned “Star history of huggingface/transformers (red) compared to other selective open source projects”. Betaworks, a_capital, Richard Socher, Greg Brockman and Kevin Durant also invested.

Delangue told AlleyWatch the same month: “We are not generating any revenue at the moment”. On choosing the lead, he said: “Given our open-source traction, it was a pretty competitive process. Despite interest from many more established Silicon Valley VCs, we decided to go with Lux, partly because of their NYC presence.”

The repository page archived the night of the announcement showed 18,958 stars, close to Reeves’s figure; 17,211 of the stars dated by then still exist. A Form D filed on 12 December 2019 reports $19,710,523 sold to 30 investors, more than the announced $15 million. The filing does not say what the difference is.

Later investors named the Hub, not the stars

Amounts are as announced or reported. The Series C valuation is from Lux’s post and the Series D valuation as reported. Stars before counts surviving stars dated before each announcement, so each figure is lower than the count shown at the time.

At the Series B, led by Addition, TechCrunch gave the library’s “42,000 stars and 10,000 forks”, but we found no statement from Addition. Lux’s post at the Series C, which it said “closed at a valuation of $2 billion”, names 10,000 companies and 100,000 models on the Hub, not stars. Salesforce Ventures, leading the Series D, wrote that “open source will shape the future of artificial intelligence” and called Hugging Face “like a GitHub for AI”. Google, Amazon, Nvidia, Intel, AMD, Qualcomm and IBM joined that round. The Series A is the only round we found where the lead investor cited GitHub stars.

Paid products came after the Series A

The first archived pricing page, from October 2020, offers a hosted inference service at “Pro 199$/month”. At the Series B, the company told TechCrunch it had been cash-flow positive in January and February 2021, that “90% of the previous round is still available on the bank account”, and that its valuation had risen fivefold. TechCrunch reported that the Series D valued the company at “more than 100 times” its annualized revenue.

Delangue has described the result two ways. In October 2024 he said the company had raised “a bit more than $500 million”, had spent “less than half of that”, and was “lucky enough to be profitable”. In a July 2026 interview quoted by TechCrunch, he said it was “close to profitability”. The announced rounds add up to $395.2 million. The Information reported in August 2026 that revenue had reached about $150 million a year. Hugging Face has not published financial statements.

People at Hugging Face open about six in ten merged pull requests

Changes to an open-source project usually arrive as pull requests, proposed edits that the maintainers review and merge. Transformers has had 21,084 pull requests merged from 3,953 people, not counting bots, and 2,671 of those people had exactly one merged. The 10 most active authors wrote 32% of them and the 50 most active wrote 58%. The top three are Yih-Dar Shieh (1,616), Sylvain Gugger (1,015) and Patrick von Platen (631).

We dated when each of the 50 most active authors worked at Hugging Face, from their own posts, company posts, archived pages, profile dates and the repository itself. For these 50, a pull request counts as Hugging Face’s when its author was at the company on the day they opened it; other authors count if they have a public link to Hugging Face. On that count, people at Hugging Face opened about half of the pull requests merged in 2020, around six in ten from 2021 to 2023, and 63% so far in 2026.

Pull requests are counted in the year they were merged, excluding bots. Stars are surviving stars dated to each year. 2026 runs to collection on 12 November. The 2019 range turns on Lysandre Debut’s first months: his Facebook profile dates his machine learning work at Hugging Face from March 2019, while his LinkedIn profile describes March to October as a growth-hacking internship before his open-source role. The 31 pull requests merged in 2018 are not shown. The 2020 range turns on Stas Bekman’s start: his post of 25 January 2021 says “Today I have joined @huggingface!”, while his own site says he was there from 2020.

That is far below Next.js, where people at Vercel opened 93% of the merged pull requests so far in 2026. Eight of the 50 most active Transformers authors never worked at Hugging Face. Five of them are Intel engineers, who opened 225 of the pull requests merged so far in 2026, about 7%. Intel invested in the Series D. We found nothing that ties the two.

Stars and merged work moved apart here too. 2023 brought 40,134 stars, twice any other year, while merged pull requests rose by about 5%.

Fewer outside pull requests are merged in 2026

Outside authors here are people with no public link to Hugging Face, not among the dated employees, and with no pull request since 2022 from a branch inside the Transformers repository, which needs write access. In 2026 they opened more pull requests than in any year measured, and a smaller share were merged.

Pull requests by outside authors, counted in the year they were opened, as of 12 November 2026. The last column covers pull requests opened from 1 January to 13 August of each year, so every year is compared at the same age.

The fall began before the maintainers wrote any rule about it. Merged outside pull requests dropped from February 2026, and the number closed without merging rose that month. On 22 March, Transformers added a “Code Agent Policy” to its pull request template: “The Transformers repo is currently being overwhelmed by a large number of PRs and issue comments written by code agents. We are currently bottlenecked by our ability to review and respond to them.” It asked new users not to submit “pure code agent PRs”. The policy says why the maintainers felt overwhelmed, but the data shows fewer outside pull requests merged; it does not show how many were written by agents.

Nvidia, a Series D investor, agreed to buy the company

Late in 2025, according to the Financial Times as reported by TechCrunch, Hugging Face turned down a $500 million investment from Nvidia at a $7 billion valuation, saying it “didn’t want a dominant investor that could sway its decisions”. Less than a year later, it agreed to sell the whole company to Nvidia. Delangue told CNBC he had approached Jensen Huang over the summer, “and a few weeks later, here we are.”

Nvidia signed the agreement on 2 September 2026. Its filing describes “an approximately $11.9 billion purchase price payable to Hugging Face stockholders” and “an equity-based retention program of up to approximately $1.0 billion for Hugging Face employees joining NVIDIA”. Huang’s post put the total at $12,930,300,000 and said “NVIDIA compute will not be required to build on or deploy through Hugging Face.” The filing says Nvidia committed to “keep Hugging Face’s platform open”, to let users “upload and download models and datasets of their choosing” and to “support other silicon vendors”.

The deal has not closed. Nvidia expects to close in the first half of 2027, subject to regulatory approvals, and Reuters reported on 28 October that the Justice Department had asked how it would affect competition in AI. On 11 November, Nvidia’s board accepted Huang’s resignation as chief executive. That filing does not mention Hugging Face.

What a small team can take from Hugging Face

These are our inferences from the record above, not claims by Hugging Face or its investors.

The pivot came from something the team already did. Hugging Face published research code while it built the app, so porting BERT was a familiar kind of work. What was new was how many people wanted it. The earlier repositories never passed a few thousand stars; this one passed 1,000 in ten days.

The company let the repository change before the pitch did. Delangue’s account puts the conversation with investors about six months after the release, and the homepage took until late summer 2019 to change. The library was renamed once in that time and again a few weeks later, each time as it took in more models.

The investor said what it saw. Lux cited stars, installs and production use, and published a star chart, in a round where the company said it had no revenue. That is unusual in this series: in the ninth Repo to Round, none of Vercel’s investors cited GitHub activity. It does not show that the stars made the round, since Reeves named installs and production use in the same sentence.

A widely used free library did not stop the company from selling. Paid inference was on sale within a year of the Series A, and the company said it was cash-flow positive by early 2021. The library stayed free and Apache-licensed throughout.

Sources and method

This article covers 2016–2026. Star figures come from GitHub’s repository star history API for huggingface/transformers, huggingface/neuralcoref, huggingface/torchMoji and huggingface/pytorch-openai-transformer-lm, collected on 12 November 2026. It returns stars still present (“surviving stars”), not historical displayed totals: the weekly data for Transformers sums to its 168,900 stars at collection. We measured removed stars by comparing archived copies of the repository page with the history: 1,316 shown on 14 November 2018 against at most 1,114 surviving, 18,958 against 17,211 on 18 December 2019, 42,269 against 39,353 on 14 March 2021, and 62,443 against 59,151 on 11 May 2022. Daily and yearly groupings use the returned week timestamps and day offsets; GitHub does not guarantee that its day boundaries align with UTC.

Pull request figures cover all 29,963 pull requests in huggingface/transformers, retrieved through GitHub’s REST API in November 2026. Of the 21,300 merged, we excluded 216 opened by GitHub app accounts, organization accounts and deleted accounts, which leaves 21,084. Commits pushed without a pull request are not counted.

For the 50 authors with the most merged pull requests, we took a start and an end date at Hugging Face from the best evidence we found, in this order: the person’s own dated statement, a company post or filing that names them, an archived page, a profile date, and the repository itself. A year with no month counts from 1 January, a month from its first day, and an end month to its last day. Engineering internships count. Of the 10,826 pull requests counted as Hugging Face’s this way, 22% rest on the author’s own statement, 5% on a company post or filing, 5% on an archived page, 66% on a profile date seen in a search result or on a third-party site, and 2% on repository evidence alone. Of the 50 authors’ 12,221 merged pull requests, 416 were opened before the author joined and 32 after the author left, and 947 are by eight authors who never worked there. Authors outside the top 50 count as Hugging Face’s if they are public members of the huggingface organization on GitHub, their profile names Hugging Face, or a merged commit carries a huggingface.co email, and that link counts for every year. Adding authors outside the top 50 who are named as reviewers in the repository’s templates changes no year by more than 1.4 points, and adding those who opened pull requests from branches inside the repository changes none by more than 3.3.

Commit counts come from the repository’s full history. Hacker News points come from the Algolia Hacker News API. Funding dates are announcement dates, and filings come from SEC EDGAR. Hugging Face is private, so revenue figures are as stated by the company or reported by the press, and marked as such above. The Financial Times report and Delangue’s July 2026 interview were read as quoted by TechCrunch.