Self-Training Agents: Hermes Agent, HF Traces, Skills, MCP & Finetuning — Merve Noyan, Hugging Face

[on hold music] Hello, everyone, and welcome to this talk in, uh, Open Agent, uh, Ecosystem, and, uh, I would like to call it Having an AI Engineer at Your Fingertips.

Um, I'm Merve, and I work in the open source team of Hugging Face. How many of you are hugging f- using Hugging Face on daily basis?

Oh, let's change that. This is not okay. Um, but first, let's talk a bit about open source and what it is. So when it comes to machine learning, open source is absolutely differential.

Basically, you have the open-weight models, um, that go in with non-commercial licenses. We call them open-weight. And then we have open source models that have, uh, commercially available licenses, such as this one from DeepSeek.

It's called, uh, uh, MIT License or Apache 2.0. And then there is, like, even more open license, uh, models that have the code open. If you have, like, agents there, the harness is open, everything is open, and this matters even more by the fact that, like, yesterday or the other day, it was revealed that the cloud, uh,

performance was going down. Uh, so if you, if you have everything in the open, nothing changes without you knowing, no performance degradation without you knowing. Everything's great. Uh, but on top of it, if you have access to the weights, you can shrink them, you can quantize them, you can fine-tune them if you feel like it.

And it's absolute guaranteed privacy for your end user because, uh, you can deploy it to edge devices, browsers without the data going somewhere else. Uh, this matters a lot, in my opinion, even more these days with the security breaches and everything.

And there was this argument maybe a few years ago that open source models aren't as good as closed.

No, no, no, no. No, this is not the case. Like you see, for instance, the latest, uh, GLM 5.1 is absolutely crushing it, and I'm actually using it in my coding setup.

Uh, the, this is the, uh, Artificial Analysis Intelligence Index, and the green ones are open models, meanwhile the black ones are the closed models. And we are-- we just catched up, and we will catch up even more with the upcoming models and stuff.

And let's go back to Hugging Face Hub. So everything is facilitated through Hugging Face Hub, all of the open releases. It's the infra layer for all of your open source, uh, workflows.

And as of now, it's hosting even more models. I should have updated the number. It's probably close to three million. A lot of datasets, spaces, and everything, but that's not all when it comes to the agentic ecosystem, and this is what we are gonna talk about today.

So when you go to the models, uh, you can filter for agentic models. Uh, they are mostly the trending ones. And there is, like, two types of models, in my opinion.

There is the v-vision LMs, and then there is the LLMs. And the vision LMs can also act as, like, a computer use agent over the screenshots. They know where to click, et cetera, which is pretty cool.

And one trend I have recently noticed is the fact that you have, uh, labs releasing their LLMs as vision, uh, with vision capabilities day zero. Like, for instance, the Gemma 4 was an omni model, and still it's an agentic model.

There is a Qwen 3.5. Uh, there is Kimi Ki- uh, Kimi K2.5. These were VLMs. So I foresee that all of these models will be, over time, uh, released day zero with vision capabilities.

And, uh, it's super easy to run this actually. Like, you can just use, like, vLLM, uh, MLX, or, like, llama.cpp, llama-server, uh, from the get go with, like, few lines of code.

Like, it used to be much more, um, friction-y, but these days, this is, uh, not a big deal.

And if you want to compare open models, we have recently launched this feature called Benchmark Datasets. So when you go to the datasets, on the left-hand side, there is like, uh, on the bottom, there is a bench-benchmark button.

You just click it, and then you can see the popular benchmarks such as SWE, uh, SWE-bench Pro or Humanity's Last Exam or AIME and others. And when you go to, for instance, SWE-bench to see, like, how your agent is, like, good in coding and stuff, uh, you see the open models ranked according to the scores.

So, like, currently, GLM 5.1 is top of the list.

So it's also easy to pick an open model these days because there's three million models out there, and it used to be a challenge to pick different models.

And if you actually want to vibe check it, Hugging Face has this ser- uh, service called Inference Providers, uh, which does routing for the best models to best providers, like all of the providers out there.

There's Groq, Cerebras, I don't know, Novita, ev-everything. And then it's super easy to compare them as well if you see, like, uh, you have the cheapest or the fastest option.

Actually, I had to truncate it, but also there is the Tool Used column, so you can actually pick one of the open source models for the agentic use case and stuff.

And going back to agents after all of these, uh, Hugging Face Hub shill, uh, Hugging Face Hub actually recently has shipped a ton of, uh, features for you to use open models with agents, agents and stuff.

And first off, like, there is the MCP server where you can plug the hub into your LLM And there is, uh, skills, uh, which allow you to even vibe train models.

Like, you just go to your agent and say, "Train Qwen two-- three point five on this data set for me," and then it just trains. Which to me is like a sci-fi at this point because it used to not exist and, like, there is so many things going on in the back and, uh, and the a-agent actually

handles them very well. And then there is the local agents, so you can run full coding agents, uh, locally from models with Hugging Face Hub because we integrate very well to them.

And coming to the first one, so basically my talk will be consisting about all of these. Uh, coming to the first one, there is the local coding agents and your options, you have, like, actually many, many options, but, like, one of my favorites is Pi because it's, like, super simple to set up.

Uh, basically you can... I, I think you can also use it with inference providers remotely, but also if you want to serve, like, a local coding agent, you can use llama.cpp to serve it, and then Pi will directly consume that.

And, uh, something very cool is also llama-agent, which is baked into llama.cpp as a binary that you can just directly execute and start a model by giving Hugging Face Hub ID.

So it's super easy as well to get a local agent running.

Uh, I will share my slides on my Twitter account after, so no need to take pictures. [chuckles] My-- one of my most favorite things these days is Hermes Agent, and I will just die on this hill.

So this is like, this is a bit one step even further to, uh, from the, uh, OpenClau by means of memory management and everything, and it's actually super easy to get started with that.

And, uh, it is you can either use it locally or with Hugging Face inference providers. So for instance, I was playing with that. Uh, like, the setup wizard does everything for you.

You just give the keys and stuff and then integrate into your Slack or WhatsApp or whatever, and you're good to go. And I absolutely recommend using this. If you want to use it with an open model, I absolutely recommend GL-GLM five point one.

For instance, I actually failed initially to integrate into Slack. I have witnesses in here, my colleague, uh, Nils is here. And, um, I asked GLM five point one to fix it, uh, with the Hermes Agent, and it's fixed on its own and it's, uh, it was a good day.

Like, uh, I, I think GLM five point one is a very good model and I cannot... I can't absolutely wait to use it with Gemma Four. But also this weekend, there was, like, on Twitter, there was a rumored, uh, MiniMax model coming up, so I will also probably try with that and share my findings.

So I absolutely recommend using Hermes Agent with the open models.

And one more thing, so basically, uh, Hugging Face Hub now has a new dataset repository type called traces, and this is basically all of your, uh, Codex, uh, Claude Code or Pi traces, they host it.

And for instance, if you go to your, um, if you pushed, uh, a trace, uh, and then you go over there, you will see in the dataset viewer, if you click on the traces column, uh, it pops up like this.

It is very nicely parsed, and you can just explore your data. And then later, if you want, you can even train a model on that, which is pretty cool in my opinion.

And, uh, if you want to push your agent traces, you, you can just upload your sessions from, uh, these, uh, paths and nothing else is needed. And we will also probably have Hermes Agent very soon for traces.

Uh, going back, if you want to use-- i-if you want more options to serve LLM behind the agent locally, so some tips and tricks in finding a good model.

You just go to Hugging Face. There is an Other tab. Under the Other tab, there is the Apps. So these apps are like LM Studio, Jeanne, um, llama.cpp, everything that is for local serving is over there.

And when you filter for them, you have the models that are supported by these, uh, by these, uh, local apps. So whatever you want to serve, we have you covered.

And when you go to the model repository, something very cool in my opinion is that on the left-hand, right-hand side, there is GGUF, uh, section. So basically GGUF, if you don't know, it's supported.

It's, it's basically comes in llama.cpp, the file, uh, format, uh, that is supported in many things like Ollama, LM Studio, everything. And you have the hardware compatibility. For instance, the Gemma Four larger model, if you quantize it to four-bit, it fits inside an L4 GPU, uh, with the twenty-four gigabyte of VRAM.

So I think this is very cool, and this is also served to, uh, MLX repositories as well. And when you go to the, again, to the model repository, if you have absolutely zero clue on how to serve this model, on top right there is Use This Model, and you have the options of the local apps that the

model is supported in. And when you click that, you see, like, only with few lines of command, uh, that you can run, you install, you get the model served, and voila.

It's very, very convenient to run the open models these days.

And lastly, supercharging your coding agents using Hugging Face skills. So there is-- we have like bunch of skills in order to get you started with training, uh, I don't know, inferring with the open models, using open models, exploring open datasets, using AI apps, everything.

And, uh, we have this thing called Hugging Face CLI skill, which allows coding agents to manage repositories, uh, run jobs, launch demos and everything. And this is how you can install it.

Uh, you can just, uh, type HF skills on Google, and you will find the, uh, commands. Uh, but we have more skills than that. So basically, this allows you to plug Hub in into your agent, like give you all of the, uh, Hugging Face Hub exploration.

But rest of the skills are super cool. There is LLM trainer skill. Basically, this is, uh, this is not only for LLMs, but also vision-language models. You can just tell the model to, okay, train this model on this dataset, and it will just kick off the job remotely, uh, on our infra or like y-locally, wherever you want.

And there is Gradio skill, which allows you to build demos, and there is Hugging Face dataset skill, which allows you to, um, explore datasets, uh, through our Dataset Viewer API, and you can install it very easily.

Again, we come with more integrations. I just put, uh, Claude and Gemini here.

So putting this into action, for instance, I asked the model, uh, to... I, I asked Claude Code to say, "Hey, can you train Qwen2 VL on LLaVA-InstructMix?" Which is like a vision-language dataset.

And it asked me a few questions. It said, "Okay, which instance would you like this to go in?" Because you have multiple options. Uh, the model actually, like in the backend, the agent actually, uh, calculates the amount of VRAM required to run, uh, fine-tune that model in a given batch size and everything.

So it handles everything for you. It just asks you a few questions, "Okay, what is your validation split?" Blah, blah. And then it just l-launches the job, which to me is absolute sci-fi still to this day as a person who have been training models since, I don't know, beginning of my career, like, uh, six, six years. [chuckles]

And you-- at the end, you just find your model on Hub. And this is not limited to LLMs and VLMs. I have recently shipped, um, skills for, for instance, training object detectors or, I don't know, s-segment anything model and everything for vision.

It handles, for instance, different bounding box types and everything. You just give the command and let it handle everything.

And going back to MCP, what do we serve? Uh, we have models, dataset, spaces, search for your task, uh, semantic search for spaces. So if you don't know spaces, it's like the App Store of AI.

You have a ton of, uh, apps over there for absolutely everything you could see. And also we have something called jobs, which allows you to kick off, uh, one of jobs that ends like, uh, if it fails or if it succeeds, and you pay for the amount of time it was up.

And also you can query these apps from MCP. Like, I'm gonna show you shortly, but it plays nicely with all of your favorite platforms.

And so for instance, in here I ask the model, "Generate image of a baklava made of yarn," and then it will call, uh, the Hugging Face space of Qwen image, which is an image generation model hosted remotely, and then it will query that and it will bring, um, the output of that.

It's works very nice. Look. But you need to turn on, there is a setting in the MCP called dynamic spaces. If you want more options of like... If you want absolutely all of the spaces, you need to turn that on, which is a bit of, b-bit experimental.

And here is some few ideas that you can use spaces MCP, uh, but you're absolutely not limited to those. And tying it all together, my colleague, Nils, has built, uh, something, I, which I found cool, so I wanted to share.

So basically, on Hugging Face Hub, there is papers and these papers, basically AI-related papers. We want people to be able to ask questions to these papers or share. Uh, but not all of the papers come with markdown, uh, which the model, which we can index and stuff.

So we OCR thirty, thirty thousand papers, uh, using Codex, Open OCR models and jobs all through prompting, which is a bit crazy. So the steps to do that is firstly, pick an OCR model that is cheap and nice and performant.

Ask the LLM to kick off a processing job and actually write the code for that and then kick it off on Hugging Face infra, and then let the skill set up the instance of hosting that model and everything without you going through the pain of the napkin math, and then profits.

So to pick an OCR model, you need to, um, you need-- you can go to All OCR Bench, which is a benchmark dataset that I have previously shown you.

The first result is Chandra OCR. But don't be fooled by this. We have just today shipped a skill that you can just ask the model, "Okay, what is the best model on OCR for fine-tuning?"

And it will also make recommendations around fine-tuning and stuff. So i-if you need like smaller models, et cetera, it will handle everything for you with this skill. So it's pretty cool.

Check it out. Um, once you pick the model, okay, we, in this case, we use Chandra. Uh, we ask model to write the script, and it did. And then the agent just does the la-napkin math for the instance and, uh, calculates the cost of the running job and everything.

And then these jobs will be... So, so basically, these jobs will be rerun. So we have recently launched this infra, uh, product called Buckets, uh, which is like, uh, S3 buckets, but much cheaper and faster, um, that you can use with, uh, mounting.

And yeah, basically, um, you can just use that. And you can get started, uh, in these links. I hope you like this talk. Thank you so much. [audience clapping] [outro jingle]