For ML engineers and developers who need to find, evaluate, run, fine-tune and deploy open models using the Hugging Face Hub, its libraries and its hosted inference options.
The curriculum is being prepared and reviewed. Enrollment is not open yet.
About this course
Hugging Face is where most open models, datasets and demos are published, and its libraries are the default way to load and fine-tune them. The stack changed substantially over the past year: the huggingface_hub 1.x release made hf the command-line tool and Xet the default storage backend, Transformers v5 went PyTorch-only and reworked quantization and loading, Inference Providers became an OpenAI-compatible router across third-party inference companies, and the ggml team behind llama.cpp joined Hugging Face in February 2026.
This course takes you from reading a model card critically to shipping an endpoint. You will manage downloads, caching and revision pinning; load models efficiently with dtype, device_map and quantization configs; serve with transformers serve or hand off to vLLM and SGLang; prepare data with Datasets and fine-tune with TRL and PEFT on one or many GPUs with Accelerate. You will compare Inference Providers, dedicated Inference Endpoints and Spaces with ZeroGPU for demos, build a small agent with smolagents and MCP, and run Hub models locally as GGUF. The final module covers organization plans, access tokens, SSO and audit, plus supply-chain hygiene: safetensors versus pickle, malware scanning and trust_remote_code. Version-specific facts are as of October 2026.
What you will be able to do
Evaluate a Hub model repository for license, gating, provenance and suitability before adoption
Download, cache and pin models and datasets reproducibly with the hf CLI and huggingface_hub
Load and run models with Transformers v5 using chat templates, appropriate precision, device placement and quantization
Fine-tune an open model with TRL and PEFT and scale training with Accelerate
Choose between Inference Providers, Inference Endpoints, Spaces and self-hosting for a workload based on cost, latency and data handling
Build a small tool-using agent with smolagents connected to an MCP server
Apply supply-chain controls including safetensors, revision pinning, scoped tokens and restrictions on trust_remote_code
Syllabus
6 modules · 17 lessons · about 12 hours
1
Module 1: The Hub as a Model Supply Chain
Find, judge, download and pin what you depend on.3 lessons
Repos, Model Cards, Licenses and Gated Models
The hf CLI and huggingface_hub: Downloads, Uploads, Tokens and the Cache
Xet Storage, safetensors and Revision Pinning for Reproducibility
2
Module 2: Transformers v5 in Practice
Load, prompt and serve models with the current library.3 lessons
Pipelines, Auto Classes and Chat Templates
Efficient Loading: dtype, device_map and Quantization Configs
transformers serve and Handing Off to vLLM or SGLang
3
Module 3: Data and Fine-Tuning
Prepare data and adapt models with the Hugging Face training stack.3 lessons
Datasets: Loading, Streaming, Processing and Dataset Cards
Supervised and Preference Fine-Tuning with TRL and PEFT LoRA
Scaling Training Across GPUs with Accelerate
4
Module 4: Hosted Inference Options
Three Hugging Face ways to run a model without managing servers.3 lessons
Inference Providers: The OpenAI-Compatible Router, Provider Selection and Billing
Inference Endpoints: Dedicated Deployments, Engines, Autoscaling and Scale-to-Zero
Spaces, Gradio and ZeroGPU for Demos and Internal Tools
5
Module 5: Agents and Local Use
Lightweight agents and taking Hub models offline.2 lessons
Building Tool-Using Agents with smolagents and MCP
GGUF on the Hub and Running Models Locally with llama.cpp
6
Module 6: Hugging Face in the Enterprise
Organization controls, security and the deployment decision.3 lessons
Team and Enterprise Organizations: SSO, Resource Groups, Fine-Grained Tokens and Audit Logs
Supply-Chain Security: Pickle Risks, Malware Scanning and trust_remote_code
Worked Decision (Hypothetical): Inference Providers, Endpoints or Self-Hosting for a Regulated Workload
Exam and certificate
The exam unlocks once you have completed every lesson. It draws 20 questions from a larger bank, spread across the modules, with the answer options shuffled for each attempt. You have 40 minutes and need 80% to pass. After a failed attempt you can try again after 24 hours, up to 3 attempts in total. Feedback is given per module, never as the answer key.
Passing earns a certificate of completion and assessment from the Wirewalk Institute of AI with a unique ID that anyone can check at institute.wirewalk.com/verify. It is not an accredited qualification or a vendor certification. How certification works.
The Wirewalk Institute of AI is independent. This course teaches publicly documented products and is not affiliated with or endorsed by Hugging Face (Hub, Transformers, Datasets, TRL, PEFT, Accelerate, Inference Providers, Inference Endpoints, Spaces, smolagents), ggml (llama.cpp), vLLM, SGLang, or PyTorch. Product and company names are trademarks of their respective owners.
A hands-on hour with ChatGPT, Claude and Gemini, told through one small business racing a deadline. Set up an account with a privacy setting you chose, decide what never goes into a personal account, write prompts that produce usable drafts, use files and web search, catch confident mistakes before they go out, and pick the assistant that fits how you work.
A practical, current tour of ChatGPT as a professional tool: choosing a plan and model, prompting well, analyzing files, researching with citations, using images and voice, organizing work in projects and memory, connecting apps and plugins, delegating multi-step work to ChatGPT Work, and knowing which admin and privacy controls matter. Ends with clear guidance on moving a workflow to the OpenAI API.
Learn Claude by solving one firm's real problem: getting through renewal season with too few analysts and without putting client data in the wrong place. Each module compares the approaches professionals actually choose between, shows a realistic failure and its fix, and ends with a Decision Lab where you make the call under pressure.