Guide · Beginner

Hugging Face: The Home of Open-Source AI

Hugging FaceBeginneractive·v5 · live·generated 35h ago

Part of these paths

Agent and Tool Ecosystem · Step 2 of 9
Enterprise Deployment Patterns · Step 7 of 12
Evaluation and Benchmarking · Step 2 of 10
Frontier Model Releases · Step 10 of 10
Inference Economics · Step 3 of 9
Long Context Evolution · Step 10 of 10
Multimodal Progress · Step 2 of 7
Open Weights Progress · Step 1 of 7
Open weights vs. the closed frontier · Step 2 of 7
Regulatory Developments · Step 8 of 9
Training Infrastructure · Step 4 of 8

TL;DRHugging Face is the platform where the open-source AI world gathers — a place to find, share, and run AI models, datasets, and tools. It started as a hub for researchers and developers, and has grown into the de facto distribution layer for nearly every major open-weights model release, while also building its own tools and even venturing into physical robotics.

Key takeaways

Hugging Face hosts landmark open-weights releases from Meta (Llama 2, 3, 3.1, 3.2, 4), Google (Gemma 3, 4, Cosmos 3), Alibaba (Qwen2.5, Qwen3), Mistral, DeepSeek, and OpenAI's GPT OSS family.
It co-created BLOOM in 2022 — a 176B-parameter multilingual model built by 1,000+ researchers — one of the first major open-access frontier models.
In early 2026 it brought GGML and llama.cpp — the key libraries that let people run AI locally on ordinary hardware — under its umbrella.
Transformers v5, a major update to its flagship open-source library, shipped in late 2025.
It acquired French robotics company Pollen Robotics in 2025, extending its open-source mission into physical robots.
Stanford's 28-trillion-pixel GPIC image dataset and DeepSeek's DSpark inference module are among the research artifacts hosted on the platform, showing its reach beyond just models.

What Hugging Face is

Hugging Face is the central meeting place for open-source AI. Picture a library crossed with a software marketplace: researchers and companies publish their AI models and datasets there, and anyone — from a solo developer to a Fortune 500 team — can download, test, and build on them. Its Transformers library (a toolkit for working with AI models in Python) is one of the most widely used pieces of software in the field.

Why it matters

Most of the AI models you hear about in the news are "closed" — you can only access them through a paid API, and you can't inspect or modify the underlying code. Hugging Face is the home of the alternative: open-weights models, where the model itself is published for anyone to use. When Meta releases a new Llama model, when Google releases Gemma, when DeepSeek publishes its latest research — they all land on Hugging Face first. That makes the platform a kind of public commons for AI development.

How it works (the basics)

You don't need to understand machine learning to benefit from Hugging Face. The Hub (its website) lets you browse thousands of models by task — text generation, image recognition, speech transcription, and more. Each model has a card explaining what it does and how to use it. Developers can pull a model into their own code with just a few lines using the Transformers library. For those who want to run AI locally on their own computer rather than in the cloud, Hugging Face now also stewards GGML and llama.cpp — the open-source tools that make running large models on ordinary laptops possible.

A platform built on open-source milestones

Hugging Face's role as the open-AI hub didn't happen overnight. A few landmarks show how it got here:

2022: Hugging Face co-created BLOOM, a 176-billion-parameter multilingual language model built collaboratively by over 1,000 researchers — one of the first open-access models to rival proprietary systems at scale.
2023: Meta's Llama 2 launched on Hugging Face, bringing millions of developers to the platform and establishing it as the default home for major open-weights releases.
2024–2025: A wave of frontier models — Meta's Llama 3, 3.1, 3.2, and Llama 4; Google's Gemma 3 and 4; Alibaba's Qwen2.5 and Qwen3 families; Mistral's various releases; and even OpenAI's GPT OSS family — all published through or announced via Hugging Face.
Late 2025: Transformers v5 shipped, a major update to the library that underpins most of the ecosystem.

Recent moves: beyond models

Hugging Face has been expanding what "open-source AI platform" means:

Robotics: In 2025 it acquired Pollen Robotics, a French company making open-source physical robots. The goal is to apply the same open-source playbook to embodied AI — robots that interact with the physical world.
Local AI infrastructure: In early 2026, GGML and llama.cpp joined Hugging Face. These are the foundational libraries that let people run large AI models on consumer hardware (a laptop, a gaming PC) rather than expensive cloud servers. Bringing them under Hugging Face's umbrella is meant to ensure their long-term maintenance and development.
Research hosting: The platform also hosts major research artifacts — like Stanford's 28-trillion-pixel GPIC image dataset and DeepSeek's DSpark inference module — showing its reach extends well beyond just model weights.

Where it's heading

The pattern across all these moves is consistent: Hugging Face wants to be the infrastructure layer for open AI, whether that means a model you run in the cloud, a tool you run on your laptop, or eventually a robot you run in your home or warehouse. As more labs — including historically closed ones like OpenAI — publish open-weights models, Hugging Face's position as the distribution hub only grows stronger.

Hugging Face as the open-AI hub

Timeline

FAQ

What exactly is Hugging Face?

Think of it as GitHub, but for AI models and datasets. It's a platform where researchers and companies publish their AI models so anyone can download, test, or build on them.

Do I need to be a researcher to use it?

No — developers, students, and companies all use it. You can browse models through a web interface, run them with a few lines of Python, or deploy them via an API.

Why do so many big companies release their models there?

Hugging Face has become the standard distribution point for open-weights AI — releasing there means instant access to a huge developer community and compatibility with widely-used tools like the Transformers library.

What is llama.cpp and why does it matter that Hugging Face acquired it?

llama.cpp (and the underlying GGML library) are the tools that let people run large AI models on ordinary laptops and desktops without expensive server hardware. Bringing them under Hugging Face's umbrella helps ensure their long-term maintenance.

Is Hugging Face only about language models?

No — it hosts vision models, speech models, multimodal models, datasets, and even robotics software since its acquisition of Pollen Robotics.

Stay current

Call Me Almanac pairs the week's AI news with guides like this one — Midweek & Sunday.

See a sample edition →

Versions

v5live35h ago
v4supersededJun 29, 2026
v3supersededJun 14, 2026
v2supersededJun 9, 2026
v1supersededJun 4, 2026

Related guides (4)

Hugging Face

Hugging Face: The Infrastructure Layer of the Open AI Ecosystem

Read asIn-depth

Hugging Face Transformers

Hugging Face Transformers: The Open-Source Library That Puts AI Models in Everyone's Hands

Read asBeginner In-depth

Anthropic

Anthropic: The AI Safety Company at the Center of the Frontier

Read asBeginner In-depth

ChatGPT

ChatGPT: The AI Assistant That Changed How the World Talks to Computers

Read asBeginner

More on Hugging Face (6)

5arXiv · cs.CL·9h ago·source ↗

MedRealMM: Real-world multimodal benchmark for Chinese online medical consultation

Researchers introduce MedRealMM, a benchmark of 5,620 real-world multimodal patient-doctor interaction cases drawn from a nationwide Chinese internet hospital, spanning 64 clinical departments. The benchmark uses a Multimodal Clinical Challenge Point (MCCP) framework to identify clinically demanding moments and evaluates 19 general-purpose and medical-specialized LLMs. Key findings show that frontier models still fall below online physician performance, particularly on safety-sensitive error avoidance, and that image information is critical for reliable clinical performance. The dataset will be publicly released on Hugging Face.

Evaluation and Benchmarking AI Safety Research Multimodal Clinical Challenge Point MedRealMM Hugging Face +1 more

7The Batch·2d ago·source ↗

DeepSeek open-sources DSpark speculative decoding module, achieving 50%+ throughput gains in production

Researchers from Peking University and DeepSeek introduced DSpark, a speculative decoding module that dynamically adjusts verification depth based on server load, achieving 57–85% faster per-user token generation and 51–52% higher total throughput compared to DeepSeek's previous production drafter. The team released checkpoints DeepSeek-V4-Pro-DSpark and DeepSeek-V4-Flash-DSpark on Hugging Face under an MIT license, with the draft module attaching to frozen target model weights. Key innovations include a parallel drafting backbone (adapted from DFlash), a Markov head for sequential token coherence correction, a calibrated confidence head, and a load-aware scheduler that trades draft length against server capacity. Results generalize across model families including Qwen3 and Gemma4.

Open Weights Progress Inference Economics Xin Cheng speculative decoding Open-PerfectBlend +7 more

4Hugging Face Blog·3d ago·source ↗

Hugging Face blog: profiling attention mechanisms in PyTorch (Part 3)

A Hugging Face blog post (Part 3 of a series) covers profiling attention operations in PyTorch, targeting practitioners who want to understand and optimize attention layer performance. The post is focused on inference/training infrastructure tooling relevant to LLM workloads. It is a practical technical guide rather than a novel research finding.

Training Infrastructure Inference Economics Hugging Face PyTorch

6Hugging Face Blog·4d ago·source ↗

Hugging Face introduces native-speed vLLM transformers modeling backend

Hugging Face announced a native-speed backend integration between vLLM and the Transformers library, enabling vLLM to use Transformers model implementations directly at native inference speed. This removes the need to maintain separate model code in vLLM, broadening model coverage and simplifying the ecosystem. The integration is significant for practitioners deploying open-weights models at scale, as it reduces friction between the two dominant open-source inference stacks.

Inference Economics Agent and Tool Ecosystem Transformers Hugging Face vLLM

5Hugging Face Blog·5d ago·source ↗

SkyPilot and Hugging Face integrate for zero-egress cloud storage across AI workloads

Hugging Face and SkyPilot have announced an integration enabling AI workloads to run on any cloud provider while storing data and model artifacts on Hugging Face Hub with zero egress fees. The partnership addresses a common pain point in multi-cloud AI infrastructure where data transfer costs create vendor lock-in. This is relevant for practitioners running distributed training or inference across cloud providers.

Training Infrastructure Inference Economics SkyPilot Hugging Face

5Hugging Face Blog·5d ago·source ↗

Hugging Face adds one-click deployment to Amazon SageMaker Studio

Hugging Face and Amazon Web Services have integrated a one-click deployment pathway from the Hugging Face Hub directly into Amazon SageMaker Studio. The integration lowers the friction for practitioners to move open-weights models from the Hub into a managed cloud inference environment. This is a partnership-level product move that deepens the Hugging Face–AWS ecosystem relationship.

Open Weights Progress Enterprise Deployment Patterns Amazon SageMaker Studio Hugging Face Amazon Web Services

Key takeaways

Hugging Face hosts landmark open-weights releases from Meta (Llama 2, 3, 3.1, 3.2, 4), Google (Gemma 3, 4, Cosmos 3), Alibaba (Qwen2.5, Qwen3), Mistral, DeepSeek, and OpenAI's GPT OSS family.

It co-created BLOOM in 2022 — a 176B-parameter multilingual model built by 1,000+ researchers — one of the first major open-access frontier models.

In early 2026 it brought GGML and llama.cpp — the key libraries that let people run AI locally on ordinary hardware — under its umbrella.

Transformers v5, a major update to its flagship open-source library, shipped in late 2025.

It acquired French robotics company Pollen Robotics in 2025, extending its open-source mission into physical robots.

Stanford's 28-trillion-pixel GPIC image dataset and DeepSeek's DSpark inference module are among the research artifacts hosted on the platform, showing its reach beyond just models.

FAQ