What is left to discover

Six chapters, six results. What follows are no longer instructions but doors: things you should know exist, so that you find them when you need them.

There is also a warning, and it is the most important paragraph in this chapter.

Agents: smolagents and Tiny Agents

Hugging Face has two very different answers to the question of how a model uses tools.

smolagents is a lean agent framework with an unusual idea: the CodeAgent lets the model write its actions as Python code, which is then executed - in a sandbox, if you do it properly. Instead of three tool calls sent one after another, the model writes a loop. Next to it stands the classic ToolCallingAgent with JSON calls. The core is deliberately kept small, under a thousand lines.

Tiny Agents is smaller still: huggingface_hub itself ships an MCP client and an agent class that comes to a few dozen lines. It connects to MCP servers, pulls their tools and hands them to a model. That is exactly where the circle closes back to chapter 05: the Gradio Space you built is such an MCP server.

When it gets serious: TGI and TEI

Two servers Hugging Face developed for production use and that you can host yourself - both as Docker containers, both OpenAI-compatible:

TGI (Text Generation Inference) serves language models with batching and the usual throughput optimisations. TEI (Text Embeddings Inference) does the same for embeddings and rerankers - the component a RAG pipeline needs between documents and database.

If that is where you are standing, the course RAG from the ground up is the sensible next step, and the building block Inference engines is the map for it.

The free courses

At huggingface.co/learnExternal - Opens in a new tab sits a whole library of courses, freely accessible, with certificates that appear in your HF profile:

• LLM Course - the broadest foundation: transformers, tokenizers, datasets, fine-tuning.
• AI Agents Course - smolagents, LlamaIndex, LangGraph; hands-on assignments submitted as Spaces.
• MCP Course - the Model Context Protocol, in cooperation with Anthropic.
• Plus courses on deep reinforcement learning, diffusion, audio, computer vision and robotics.

They are text-based and have no deadlines. Anyone coming out of this course who wants to go deeper belongs in the LLM Course; anyone who got stuck on chapter 05 belongs in the Agents Course.

Papers and collections

Papers links arXiv publications with the models, datasets and Spaces that belong to them. It is the shortest road from "I read about that" to "I can try it out".

Collections are sets of models, datasets and Spaces somebody has assembled around a topic. Creating your own is the simplest way to remember your own research - and considerably more usable than thirty browser tabs.

And now the warning

The Open LLM Leaderboard was for years the best-known page on the Hub: a ranking of open models, used monthly by hundreds of thousands of people.

It was archived in March 2025. It is still there, it still opens, it still looks the way it did - but it is no longer updated. What you see there is a still image from the spring of 2025.

The reasons for archiving it are instructive, incidentally: the benchmarks were saturated (too many models crowded at the top with nearly identical scores), the compute cost was considerable, and evaluation shifted towards human preference. That is not a failure of the list but a hint at how fast evaluation methods age.

What to use instead:

• Arena-style comparisons, where people choose between two answers.
• Task-specific rankings - the MTEB leaderboard for embeddings, comparisons such as EuroEval for European languages.
• Your own test. Twenty real examples from your use case say more than any general ranking. The building block Benchmarks explains why.

The milestone

No code this time, but a file. Write down three things:

1. Which door you open next - agents, production servers, fine-tuning, or one of the courses. One, not four.
2. Why, in one sentence, with reference to a real task.
3. A bookmark on a ranking or comparison page that is still maintained - with the date on which you checked that it is current.

Point 3 is the real test of this chapter. Anyone who can answer it has understood the lesson: on a platform that moults every few months, knowing what is out of date is worth as much as knowing what exists.