HuggingFace

Hugging Face develops infrastructure for sharing and deploying AI: the Transformers library provides a unified API for pretrained models, while the Hub brings together models, datasets and application demos. Its own research includes multimodal models and small language models, alongside collaborative work such as BigScience’s BLOOM and an expansion into robotics through LeRobot and Pollen Robotics.[3][2][1][4][5][6]

On September 2, 2026, NVIDIA entered a definitive agreement to acquire Hugging Face, Inc. The transaction included approximately $11.9 billion payable to stockholders, subject to adjustments, and an equity-based employee retention program of up to approximately $1.0 billion. NVIDIA expected closing in the first half of 2027, subject to customary conditions and required regulatory approvals. Jensen Huang’s September 3 announcement separately gave a headline figure of $12,930,300,000.[9][8]

NVIDIA committed to keeping the platform open, including uploads and downloads of users’ chosen models and datasets and support for other silicon vendors. Huang said NVIDIA compute would not be required and that multi-cloud and multi-accelerator development and deployment would continue. These were commitments accompanying the proposed acquisition.[9][8]

The platform’s earlier history begins with a consumer chatbot. Clément Delangue, Julien Chaumond and Thomas Wolf founded Hugging Face in 2016, initially as an AI chatbot company. In its 2023 retrospective, Salesforce Ventures described the founders’ shift toward the natural language processing models powering that chatbot, followed by the company’s open-source work in 2018 and 2019.[2]

Transformers made different Transformer architectures available through a unified API, backed by a curated collection of pretrained community models. This gave practitioners a common interface while allowing researchers to extend the library. The Transformers paper, first submitted on October 9, 2019, credited Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clément Delangue and a wider group of coauthors.[3]

By August 2023, Hugging Face’s products covered sharing, training and deployment. The Hub hosted models, datasets and demos; Spaces let users build and share machine-learning applications. Inference endpoints deployed models into production, AutoTrain supported scalable low-code training, and private Hubs provided internal collaboration and model hosting. Models hosted on the platform included work from other organizations, such as Meta’s LLaMA 2 and Salesforce’s BLIP.[2]

Hugging Face released BLOOM through the BigScience collaboration on July 12, 2022. The release described a 176-billion-parameter model capable of generating text in 46 natural languages and 13 programming languages. It reported participation by more than 1,000 researchers from over 70 countries and more than 250 institutions, with a final 117-day training run on France’s Jean Zay supercomputer. CNRS and GENCI supplied a compute grant estimated at €3 million.[4]

BLOOM could be downloaded, studied and used locally or through a cloud provider subject to its Responsible AI License, developed within BigScience. The release also made intermediary checkpoints and optimizer states available, opening access to training artifacts as well as the finished model.[4]

In 2023, Hugging Face’s M4 multimodal research team pursued an independent open-access reproduction of DeepMind’s Flamingo. The effort began in March and released IDEFICS in August, in 9B and 80B sizes. IDEFICS connected Llama v1 language modeling with a LAION CLIP vision model, accepting interleaved images and text for tasks such as image description, visual questions and creative text generation. The January 2024 account featured M4 researchers Hugo Laurençon and Leo Tronchon and identified Victor Sanh as an IDEFICS coauthor.[1]

The team created OBELICS to supply interleaved image-text documents for training, preserving richer document context than isolated image-caption pairs. The January 2024 account described 115 billion text tokens, 141 million English documents and 353 million images, curated and filtered from Common Crawl dumps spanning February 2020 through February 2023. It also described an engineering decision to retain the LAION CLIP encoder after an experiment with the larger EVA-CLIP encoder encountered training instability.[1]

Hugging Face began LeRobot in 2024 under Rémi Cadène, who had worked on Tesla’s Optimus project. Its robotics work included collaboration with Pollen Robotics on Reachy 2 in June 2024, with The Robot Studio on the roughly $100 SO-100 arm in October, and with NVIDIA on data collection, training and verification workflows in November.[6]

On April 14, 2025, Hugging Face announced its acquisition of Pollen Robotics, calling it its fifth acquisition after additions including Gradio and XetHub. Pollen, founded in 2016 by former Inria Flowers researchers including Matthieu Lapeyre and Pierre Rouanet, brought its own history of open-source robot hardware. Hugging Face offered Pollen’s Reachy 2, an open-source, VR-compatible humanoid for research, education and embodied AI experiments, for $70,000 at announcement.[6]

On July 8, 2025, Hugging Face released SmolLM3, a 3B-parameter model with reasoning and non-reasoning modes and support for English, French, Spanish, German, Italian and Portuguese. Its published recipe mixed web, mathematics and code data across three pretraining stages. Architectural changes included four-group grouped-query attention to reduce the inference KV cache and removal of rotary position embeddings from every fourth layer. The team reported training on 384 H100 GPUs for 24 days, using nanotron for training, datatrove for data processing and lighteval for evaluation.[5]

SmolLM3’s long-context adaptation extended training sequences to 64k tokens, with YaRN extrapolation supporting inference up to 128k. The team found that subsequent reasoning training degraded long-context performance. It used MergeKit to combine a soup of Anchored Preference Optimization checkpoints with a strong long-context mid-training checkpoint, weighted 0.9 and 0.1 respectively. The team reported that this recovered the base model’s RULER score on contexts up to 128k tokens.[5]

Hugging Face announced Community Evals on February 4, 2026, in beta. Benchmark dataset repositories could host leaderboards that aggregated scores from model repositories. Benchmark specifications lived in eval.yaml, based on the Inspect AI format, while model scores lived in .eval_results/*.yaml. Any user could submit scores through a pull request; submissions appeared as community results before model-author acceptance, although authors could close score pull requests and hide results.[7]

Community Evals made evaluation specifications, reported scores and their Git history visible. Its authors explicitly cautioned that the feature would not solve benchmark saturation, the gap between benchmarks and real-world performance, or training on test sets.[7]

Salesforce Ventures announced on August 24, 2023, that it was leading Hugging Face’s Series D, reporting more than 2 million users and more than 1 million Hub repositories encompassing models, datasets and demos. In April 2025, Hugging Face reported more than $395 million raised cumulatively, more than 7 million users and more than 500,000 Spaces.[2][6]

In the September 3, 2026 acquisition announcement, Huang reported more than 18 million developers, researchers and creators, sharing more than 3 million models, 500,000 datasets and 1 million applications. He also reported more than 200,000 companies using Hugging Face to discover, evaluate, customize and deploy AI. These were platform-use figures, rather than counts of paying customers.[8]