Big Tech pumps $235M into AI model depot Hugging Face

Hugging Face, which maintains a huge repository of open and non-open source AI models and training data, just closed a $235 million funding round backed by top tech players and a Hollywood heavyweight.

The latest series-D round, led by Salesforce Ventures, catapults Hugging Face’s valuation to a whopping $4 billion. Other notable investors include: Google, Amazon, Nvidia, AMD, Intel, IBM, Qualcomm, and Sound Ventures, an investment firm led by actor Ashton Kutcher, a representative from Hugging Face confirmed to The Register. 

There’s an interest in openly available AI models as developers weigh up whether it’s worth using proprietary systems. Being able to train and fine-tune an open model on custom data allows organizations to exert greater control over the technology before embedded it in their products and services, if the license allows, it’s felt.

Openly available models may be easier to shape and maintain, and be potentially cheaper to run in-house, versus black-box models available through some pay-per-word cloud API. Those closed-source solutions may also regress over time, or be subject to inconsistent content moderation and other changes that developers may not want to have to deal with.

If you have a model at your fingertips and can do with it as you please with it, license conditions permitting, you won’t be subject to the whims of some large, faceless machine-learning tech provider.

Hugging Face hosts hundreds of thousands of models and datasets on its platform provided by other developers, allowing users to access the code and resources they need to build their own applications. These materials are available under a mix of licences, from open source ones like the MIT license to research-friendly non-commercial agreements.

It hosts some of the most popular models out there, such as Meta’s LLaMa series or Stability AI’s Stable Diffusion, as well as ones built by teams of developers including Falcon or Vicuna. Last year, Hugging Face also trained and released its own open-access large language model BLOOM.

• Meta can call Llama 2 open source as much as it likes, but that doesn’t mean it is
• Small custom AI models are cheap to train and can keep data private, says startup
• Open source AI makes modern PCs relevant, and subscriptions seem shabby

Open source and freeware-slash-shareware machine-learning models have caught the attention of Google or Amazon: while those kinds of Big Tech players offer cloud-hosted closed-source, proprietary models, it’s in their interests to support open models, which can be run in public clouds, resulting in more sales for those providers.

Salesforce, for one, has also invested in private large language model providers Cohere and Anthropic, and is now backing Hugging Face.

“We believe the future of AI will be driven by a thriving open ecosystem of researchers, developers, enterprises, and enthusiasts collaborating together. Hugging Face sits at the center of the open AI movement and has become the most important community for ML researchers and practitioners,” the software-as-a-service titan’s investment arm, Salesforce Ventures, said in a statement. 

Hugging Face makes money by charging developers and enterprise customers for features that make it easier for them to train and run AI models at scale, including providing extra tools, compute power, or security. 

CEO Clement Delangue welcomed the investment, natch, and said: “These partners alone already shared over 1,000 open models and datasets and have over 10,000 users on Hugging Face. It takes a village to democratize good machine learning thanks to open-source and we’re just getting started.”

The biz is based in Brooklyn, New York, and employs about 170 people, so far. ®