Overview
Hugging Face is the world's largest AI open-source community and model hosting platform, known as the "GitHub of AI." It has more than 18 million developers and hosts 3M+ open-source models, 500K+ datasets and 1M+ AI app demos (Spaces). Almost every major open-source AI model is released on Hugging Face.
Hugging Face is not just a model repository; it also provides core open-source libraries like Transformers, Inference API, training platforms (AutoTrain) and other development toolchains, making it essential for AI developers.
**🔥 Acquired by NVIDIA on 2026/9/3**: NVIDIA agreed to acquire Hugging Face for about **$12.93 billion**, comprising roughly $11.9 billion in cash plus up to $1 billion in employee equity retention, with the deal expected to close in the first half of 2027 pending regulatory approval. Jensen Huang pledged that Hugging Face will operate as an independent business unit, remain a neutral open platform, and will not require NVIDIA compute. Before that, in May 2026, Hugging Face had closed a **$500 million Series D at an $8 billion post-money valuation**. Spaces now also integrates **MCP (Model Context Protocol)** end to end: any AI app or robot (such as Reachy Mini) can call public tools on Spaces (speech, vision, generation, external models) over MCP, upgrading Hugging Face from a model repository to an AI tool marketplace.
Key Features
- 2 Million+ Open-Source Models: The world's largest AI model library, covering NLP, CV, audio, multimodal, and other fields
- Spaces Application Hosting: Host and share AI application demos, one-click deployment of Gradio/Streamlit apps, supporting MCP protocol and new features like real-time voice
- Transformers Library: The most popular AI model library, load and use any model with just a few lines of code
- Inference API: Call models on the platform via API without deploying them yourself
- Datasets: 100,000+ open-source datasets, convenient for model training and evaluation
- AutoTrain: No-code model fine-tuning tool, upload data to train
Use Cases
- AI researchers and developers finding and using open-source models
- Enterprise teams evaluating and deploying open-source AI solutions
- AI learners experimenting and practicing with models
- Sharing open-source models and community collaboration
- Quickly building AI application prototypes
Pros
- Largest model library: almost all important open-source models are here
- Completely free: basic features and a large amount of resources are free to use
- Transformers library ecosystem: the standard library for Python AI development
- Active community: good atmosphere for paper reproduction, discussion, and collaboration
- Convenient Spaces: quickly share AI demos, supporting MCP protocol and new features like real-time voice
Pricing
Hugging Face's basic features are completely free (model downloads, Spaces hosting, community features). Pro $9/month (more Spaces resources, private models, early access to new features). Enterprise Hub pricing on request (enterprise-grade security, SSO, auditing). Inference Endpoints billed based on GPU usage.
Summary
Hugging Face is an essential platform for AI developers: whether you want to find models, share projects or learn AI, you can hardly avoid it. For developers its status is like GitHub for software development. After NVIDIA's $12.93 billion acquisition, the platform has pledged to stay neutral and open, so the developer ecosystem is unaffected in the near term. Non-technical users may not need it, but if you work on AI development, not having a Hugging Face account is hard to justify.
Version History
- NVIDIA to acquire Hugging Face for $12.93B, pledging to keep the platform open (2026-09-03): NVIDIA announced a definitive agreement to acquire Hugging Face for about $12.93 billion (roughly RMB 87.09 billion), comprising about $11.9 billion in cash and up to $1 billion in employee equity retention, expected to close in the first half of 2027 pending regulatory approval. Jensen Huang said Hugging Face will operate as an independent business unit and remain a neutral, open platform: developers choose their own models, frameworks, cloud providers and compute platforms, NVIDIA compute is not a requirement, and multi-cloud and multi-accelerator deployment continue to be supported. The platform now has over 18 million developers, hosting 3M+ models, 500K+ datasets and 1M+ apps, used by more than 200,000 companies.
- Hugging Face ships @huggingface/kernels with 207 WebGPU kernels for the browser (2026-09-01): Hugging Face WebAI team releases the @huggingface/kernels library and 207 WebGPU kernels hosted as independent repositories on the Hub (Apache-2.0), each with a manifest, correctness tests, benchmark cases, and WGSL shader templates.
- Measuring benchmark optimization in speech recognition: new Hugging Face tests expose ASR benchmark gaming (2026-08-21): Hugging Face's latest research introduces three tests to quantify the phenomenon of benchmark optimization (benchmaxxing) in speech recognition. An evaluation of 11 open-source ASR models shows that multiple high-scoring systems reproduce erroneous transcriptions from the VoxPopuli and LibriSpeech benchmarks, even when the audio content contradicts them. Some models even rely on acoustic cues to identify the benchmark source, causing their scores to overestimate true transcription capability.
- Hugging Face releases LFM2.5 DSpark draft models, up to 3.18x faster inference (2026-08-20): Hugging Face released the DSpark draft model checkpoints for three models in the LFM2.5 series, achieving up to 3.18x GPU throughput improvement and 2.87x on-device improvement through speculative decoding without changing output quality. The draft model has approximately 300M parameters, and LFM2.5-2.6B function calling latency is reduced by an average of 57%. It is open-sourced with support for llama.cpp and SGLang.
- SGLang announces Day-0 support for NVIDIA Nemotron 3.5 Lightning (2026-08-11): SGLang announces Day-0 support for NVIDIA Nemotron 3.5 Lightning, an open-source model with a Mixture-of-Experts architecture of 30B total parameters and 3B activated parameters, supporting up to 1M token context. BF16 and NVFP4 weights can be downloaded from Hugging Face. The model supports three speculative decoding techniques: MTP, DFlash, and DSpark, and can be integrated into agent workflows via an OpenAI-compatible API.
- FeyNoBg released: open-source background removal model reaching SOTA on four benchmarks (2026-07-28): Feyn Labs has launched FeyNoBg, a state-of-the-art model for automatic background removal. It achieves the best S-measure scores on four out of eight benchmarks, while on the remaining four, it trails the leader by less than 2%. The model is based on the BiRefNet architecture, with parameters expanded from 222M to 263M. Additionally, the training library NoBg has been open-sourced, with the model and code available on Hugging Face and GitHub, respectively.
- Microsoft open-sources the MagenticLite models (2026-07-22): The models of MagenticLite are now fully open-sourced. The MagenticBrain and Fara 1.5, previously available on Microsoft Foundry, now have their weights released on Hugging Face. The application, testing tools, and every model in the stack are now fully open.
- Shanghai Scientific Intelligence Research Institute opens an 11B open-science multimodal foundation model (2026-07-20): The Shanghai Scientific Intelligence Research Institute released the open science multimodal foundation model "Shenzhen," with a total of approximately 11 billion parameters, capable of processing six types of data: DNA, RNA, proteins, small molecules, Earth systems, and medical images. Among 20 tasks in biological sequences, the model achieved the best results in 9 tasks; the average Dice score for medical image segmentation was 91.20, the best among 7 evaluated methods. Model weights and code have been open-sourced on Xinghe Qizhi, Hugging Face, and GitHub.