Hugging Face

Hugging Face is an American AI company and open-source platform that has become the central repository and community hub for the global machine learning ecosystem. Founded in 2016 and headquartered in New York City with offices in Paris, it initially launched as a consumer chatbot application before pivoting to become the infrastructure layer of open AI development. Often described as the "GitHub of machine learning," Hugging Face hosts models, datasets, and deployable applications across virtually every AI domain, and maintains widely adopted open-source libraries including the Transformers, Diffusers, and Datasets packages. The platform has achieved a dominant position in open-source AI by offering the distribution infrastructure, tooling, and community that enables researchers and practitioners to share, discover, and build upon each other's work. As of 2025, Hugging Face reports 13 million registered users, over two million publicly available models, and more than 500,000 public datasets, with between 1,000 and 2,000 new models uploaded daily.^[1]

The Hugging Face Hub

The Hub is the core platform product: a version-controlled repository for model weights, configuration files, training scripts, and documentation. Each model repository includes a model card—a standardised document describing the model's intended use, training data, performance benchmarks, limitations, and ethical considerations. This documentation standard has influenced broader industry practice around model transparency. The Hub supports all major model formats and frameworks, including PyTorch, TensorFlow, JAX, ONNX, and GGUF, and integrates directly with popular training frameworks. Authentication, access controls, and private repositories allow commercial users and institutions to host proprietary or restricted models within the same infrastructure.

Transformers Library

The Transformers Python library, first released in 2019, provides a unified API for loading and running thousands of pre-trained models across natural language processing, computer vision, audio, and multimodal tasks. Its design abstracts away the differences between model architectures, allowing users to swap between BERT, GPT-2, LLaMA, Mistral, Falcon, and hundreds of other architectures with minimal code changes. The library has become the de-facto standard for applied NLP and has been cited in tens of thousands of academic papers.^[2] Supporting libraries include Diffusers (for stable diffusion and other image-generation models), Datasets (a unified data loading and processing API), Evaluate (standardised model evaluation), PEFT (parameter-efficient fine-tuning methods including LoRA), and TRL (transformer reinforcement learning, for RLHF and alignment training).

Spaces

Spaces is Hugging Face's application hosting platform, which allows developers to deploy interactive machine learning demos built with Gradio, Streamlit, or static HTML. With over 500,000 live applications, it has become the largest directory of AI demos and tools, enabling researchers to share working prototypes alongside their model releases. Spaces can run on free CPU instances or be upgraded to GPU-backed compute for more demanding models.

Business Model and Commercial Products

Hugging Face's revenue reached an estimated USD 130 million in 2024, up from USD 70 million in 2023, serving approximately 50,000 enterprise customers.^[1] Commercial products include the Inference API (hosted model inference), Inference Endpoints (dedicated, private model deployments on cloud infrastructure), and AutoTrain (no-code fine-tuning for custom datasets). The company raised a USD 235 million Series D round in 2023 at a valuation of USD 4.5 billion, with investors including Google, Amazon, Nvidia, Salesforce, and Intel.

Ecosystem and Community

Hugging Face's open-source community has grown to include individual researchers, academic labs, national AI institutes, and major technology companies. Meta AI, Mistral AI, Google DeepMind, Microsoft, and Stability AI all publish models on the Hub as part of their open-source strategies. In 2025, NVIDIA deposited over 650 models on the platform, and Meta co-launched a hub for agentic environments. Robotics datasets—a fast-growing research area—experienced a 24-fold increase in uploads, reflecting the platform's expanding scope beyond language.^[1] Hugging Face has become an important resource for the Malaysian AI community, particularly for researchers and developers working on Bahasa Malaysia and multilingual Southeast Asian language tasks. The platform hosts several Malay-language datasets and fine-tuned models contributed by Malaysian research institutions including Universiti Malaya (UM), Universiti Teknologi Malaysia (UTM), and independent contributors. These include sentiment analysis datasets for Malay social media text, Malay question-answering benchmarks, and code-switching corpora that reflect the mixed-language communication patterns common in Malaysian digital contexts. MDEC (Malaysia Digital Economy Corporation) has featured Hugging Face in its AI ecosystem mapping reports as a key upstream resource for Malaysian startups building AI-powered products. Local companies developing NLP applications for Bahasa Malaysia typically use the Hugging Face Transformers library as their core development toolkit, fine-tuning multilingual base models such as mBERT, XLM-R, or Mistral on local datasets. Malaysia's AI talent development programmes have incorporated Hugging Face tools extensively. HRD Corp-approved AI training providers use the platform's free tier and Transformers library as the practical foundation for instruction in NLP, computer vision, and model deployment. The Spaces feature allows trainees to demonstrate working prototypes without needing to configure their own inference infrastructure—a meaningful advantage given Malaysia's growing but still developing cloud engineering talent base. From a commercial standpoint, Malaysian fintech and legaltech startups accessing the Hugging Face Inference API benefit from predictable costs and reduced operational overhead compared with self-hosting. The Personal Data Protection Act (PDPA) has led some Malaysian enterprises to favour Hugging Face Inference Endpoints—dedicated private deployments within AWS or Azure regions that can be configured for data residency in Southeast Asia—over the shared Inference API for sensitive workloads. As Malaysian cloud infrastructure expands (with AWS, Google, and Microsoft all establishing data centres in Malaysia by 2025), running Hugging Face-hosted inference workloads with Malaysian data residency is increasingly feasible.

See Also

References

1. Callin. (2025). Hugging Face: The Comprehensive Guide to AI's Most Dynamic Platform in 2025. https://callin.io/hugging-face/
2. Hugging Face. (2024). What is the Hugging Face Community Building? https://huggingface.co/blog/evijit/hf-hub-ecosystem-overview
3. Ultralytics. (2025). What is Hugging Face? AI Model Hub. https://www.ultralytics.com/glossary/hugging-face