Where the AI world stores its models — a tour through the largest repository.
Architectures 12 min Beginner June 22, 2026
You want to use an AI model beyond ChatGPT's chat window. But where do you actually find models? How do you tell which one is good? And what does it mean when a model is labeled "open source"?
Hugging Face is the central platform for the open AI ecosystem — the place where models, datasets, and demos converge. This chapter shows you how to navigate it, read a Model Card, and critically evaluate leaderboard scores.
The Platform — Your Map to the AI Model World
Hugging Face organizes the open AI ecosystem around three pillars. Each pillar has a clear function — together they form the infrastructure on which the entire open-source AI ecosystem is built.
Model Hub A searchable repository hosting over one million pre-trained models — language models, image generators, audio and video models.
Datasets Hub Hundreds of thousands of curated datasets for training, fine-tuning, and evaluating models.
Spaces Interactive web demos (Gradio/Streamlit) where you can try models directly in the browser — no code, GPU, or setup required.
1M+
Models in the Model Hub Freely searchable and downloadable
Think of Hugging Face as a massive public library. The Model Hub is the book collection — over a million titles, organized by category, language, and popularity. The Datasets Hub is the archive room containing the raw research material that the books were written from. And Spaces are the reading rooms where you can open any book and try it out interactively, without taking it home.
Just like a library, the most popular books are easy to find (sorted by downloads and likes), but a niche book that perfectly fits your topic might be more useful than the bestseller.
Example: You type "llama 3" into the Model Hub search bar. The top result is meta-llama/Llama-3.1-8B-Instruct — an official Meta repository.
The page shows millions of downloads, thousands of likes, the full Model Card with architecture details and benchmark scores, and a "Use this model" button that generates the Python code to load it in three lines. In the Spaces tab, you'll find dozens of community-built demos.
Misconception: The models on Hugging Face come from Hugging Face
Hugging Face is primarily a hosting and infrastructure platform, and only marginally a model developer (e.g. SmolLM). The vast majority of models come from organizations like Meta (Llama), Google (Gemma), Microsoft (Phi), Mistral, and thousands of independent researchers. Hugging Face provides the distribution channel, the standardized documentation format, and the community tools — similar to how GitHub does not write software but enables it to be shared.
Model Cards & Weights — What's Inside a Model?
When you open a model repository on Hugging Face, you find two things: a Model Card (structured documentation) and the model weights (the actual learned parameters).
Model Card
AnalogyDefinition
A Model Card is like a package insert for a medication: It describes what the model is designed for, how it was trained, what limitations it has, and under what conditions you may use it.
Analogy:
A Model Card is like a package insert for a medication: It describes what the model is designed for, how it was trained, what limitations it has, and under what conditions you may use it.
Definition:
A standardized documentation format (proposed in 2019 by Margaret Mitchell et al.) containing model description, training data, evaluation results, known limitations, and licensing terms in structured form.
The weights are the gigantic matrices of numerical values that the model learned during training. They are stored as large binary files (typically in safetensors format), and their size scales directly with parameter count.
Open Weights
AnalogyDefinition
Open weights are like a restaurant selling you their signature dish to take home: You can eat it, reheat it, and add toppings — but you don't get the recipe, the ingredient list, or the kitchen equipment specifications.
Analogy:
Open weights are like a restaurant selling you their signature dish to take home: You can eat it, reheat it, and add toppings — but you don't get the recipe, the ingredient list, or the kitchen equipment specifications.
Definition:
The finished model parameters are publicly available, but training code and training data remain proprietary. Most models labeled "open source" (like Meta's Llama) are actually open-weights models.
True open source is the complete recipe book: ingredients, preparation steps, temperatures, and plating instructions — so you could recreate the dish from scratch in your own kitchen. Example: OLMo publishes everything — training code, dataset, and weights.
Open Weights (e.g. Llama)
Weights: yes | Training code: no | Training data: no | License: restricted (700M user threshold)
Open Source (e.g. OLMo)
Weights: yes | Training code: yes | Training data: yes | License: fully open
~16 GB
Storage (8B model) Approx. 16 GB at full 16-bit precision — more parameters mean larger files
Example: The Model Card for mistralai/Mistral-7B-Instruct-v0.3 tells you: 7 billion parameters, a vocabulary of 32,768 tokens, Apache 2.0 license (full commercial freedom). According to Mistral, it matches or beats Llama 2 13B (a model nearly twice its size) on several tasks. The Limitations section points out that the model has no moderation mechanisms.
Caution: Available does not mean license-free
Even when weights are publicly available, licenses vary enormously. Llama 3.1 comes with the "Llama 3.1 Community License" that restricts usage above 700 million monthly users. Some models are released exclusively for non-commercial research. Always check the license field in the Model Card before building a product.
Leaderboards — How Models Are Ranked (and Why Rankings Lie)
With over a million models available, the community needs systematic comparison methods. Two dominant approaches have emerged: automated benchmarks and human evaluation.
Automated Benchmarks
e.g. the former Open LLM Leaderboard (retired in 2025): MMLU, ARC, HellaSwag, GSM8K — standardized, scalable, but gameable
Chatbot Arena (now: Arena)
Elo-style rating: Blind human evaluation — practical, but subjective and slow
Automated benchmarks are like car magazine comparison tables: 0-to-60 acceleration, fuel efficiency, trunk volume — precise numbers that rank every car on paper. The Chatbot Arena is the test drive: you sit in two cars and decide which one you actually prefer for your daily commute. A car that tops the magazine charts might feel terrible to drive.
Deep Dive: What do the key benchmarks measure?
MMLU (Massive Multitask Language Understanding): 57 knowledge domains from astronomy to law — tests broad factual knowledge.
HellaSwag: Complete sentences meaningfully — tests language understanding and common sense.
GSM8K: Grade-school math word problems — tests step-by-step calculation and problem solving.
Benchmark Gaming: When Scores Lie
A model developer notices that MMLU questions follow predictable patterns. They add MMLU-style questions to their training data — not the exact questions, but structurally similar ones. Their model jumps 5 points on the leaderboard. Users download it excitedly — only to find it performs no better (or even worse) on their actual tasks.
Misconception: The top model is the best for my project
Leaderboards measure average performance across diverse tasks. A 70-billion-parameter model ranking first overall may be dramatically outperformed by a fine-tuned 7-billion-parameter model on your specific task (e.g., German legal text summarization). The right model depends on your use case, your hardware, and your latency requirements.
Interactive: Comparing Model Cards
Here the models from this chapter sit side by side, the way their model cards describe them: size, license, and what was actually released. The table estimates storage like the 8B model above: roughly 2 GB per billion parameters. Sort by a column or click a model.
Key Takeaways
Hugging Face is the central infrastructure of the open AI ecosystem with three pillars: Model Hub, Datasets Hub, and Spaces.
A Model Card tells you what a model can do, how it was trained, and under what conditions you may use it.
"Open weights" is not the same as "open source" — always check what was actually published.
Leaderboard scores are a starting point for comparison, not a final verdict — the right model depends on your specific task.
You now know where AI models live and how to evaluate them. In the next article, we bring a model to your own computer: With Ollama and LM Studio, you can run language models locally — no cloud, no subscription.
Quiz: Hugging Face
Question 1 / 5
Not completed
What is the primary function of "Spaces" on Hugging Face?
1. What is the primary function of "Spaces" on Hugging Face?
☐ A) Storing training datasets for model development
☐ B) Hosting interactive web demos where users can try models in the browser
☐ C) Providing GPU compute for training new models
☐ D) Managing Git repositories for model version control
2. A model on Hugging Face is labeled "open source" and publishes its weights under Apache 2.0 — but training code and training data are not published. What is the accurate term?
☐ A) Fully open source
☐ B) Open weights
☐ C) Proprietary
☐ D) Freeware
3. You need a model for summarizing German legal documents. Model A ranks #1 on a leaderboard (highest MMLU score). Model B ranks #15 but was fine-tuned specifically on German legal texts. Which should you evaluate first?
☐ A) Model A — the leaderboard proves it is better at all tasks
☐ B) Model B — task-specific fine-tuning typically outperforms generic scores in specialized domains
☐ C) Model A — more parameters always mean better results
☐ D) Neither — for legal tasks you should only use proprietary models
4. A Model Card lists the license as "Llama 3.1 Community License". Your startup plans to integrate this model into an app with 50,000 users. What should you check in the license?
☐ A) Nothing — Community License means it is completely free
☐ B) Whether there are user-count thresholds that trigger commercial restrictions
☐ C) Whether Meta approves each commercial use case individually
☐ D) Whether the model may only run on Meta's servers
5. A new model jumps 8 points on the MMLU benchmark overnight but shows no improvement in real user interactions on the Chatbot Arena. What is the most likely explanation?
☐ A) The Chatbot Arena is broken and needs recalibration
☐ B) The model was likely optimized on MMLU-style patterns (benchmark gaming) without genuine capability improvement
☐ C) MMLU and Chatbot Arena measure exactly the same thing, so the scores should be identical
☐ D) The model needs more parameters to translate MMLU gains into chat quality
Answer Key: 1) B · 2) B · 3) B · 4) B · 5) B
Checkpoint
What three main areas does Hugging Face offer and what is each one for?
Why is a model like Llama 3 "open weights" but not "open source"?
Why shouldn't you blindly trust the #1 spot on a leaderboard?