16 self-contained Jupyter notebooks. Open one, choose Run All, and a complete AI application builds itself on your own GPU — no cloud account, and the model runs on your own hardware.
Each demo checks your environment, installs what it needs, starts the model, and opens an interactive app.
Important
Evaluation and demonstration use only. These notebooks are sample code to help you evaluate AI workloads on Intel® Arc™ Pro GPUs. They are not production software and are provided "as is", without warranty of any kind. You are responsible for how you run them, for any data you provide, and for validating any results before relying on them. See DISCLAIMER.md.
git clone https://github.com/intel/intel-xpu-ai-demos.git
cd intel-xpu-ai-demos
python -m pip install -r requirements.txt # JupyterLab, if you don't have it
jupyter labThen open any demo's notebook and choose Run All. New to this? Start with Hello World — vLLM Edition — it is the shortest path from nothing to a model answering a question on your GPU.
What you need: an Intel® Arc™ Pro B50, B60 or B70 with its GPU driver installed, Ubuntu 24.04 / 25.10 / 26.04, Docker, and an internet connection for the first run. Full details in docs/SETUP.md.
First contact with model serving on an Intel GPU — begin with one of these.
| Demo | What it does | Runs on |
|---|---|---|
| ⚡ Hello World — vLLM Edition | The simplest way to see production AI serving on an Intel GPU: launch vLLM in Docker, then chat with a live model over the OpenAI-compatible API. | vLLM in Docker |
| 🐉 Hello World — SGLang Edition | The same first-steps demo on a different engine: SGLang with RadixAttention prefix caching, in Docker, over the OpenAI-compatible API. | SGLang in Docker |
Several modalities, or several models, working together in one pipeline.
| Demo | What it does | Runs on |
|---|---|---|
| 💎 AI Studio with Gemma 4 | Run Google's Gemma 4 multimodal models on Intel GPUs — text, images, documents, OCR, charts, screenshots, audio and video in one notebook. | vLLM in Docker |
| 🎬 Multi-Model Orchestration | One director model writes briefs for four specialist models, which then work in parallel on separate GPU tiles and get reviewed at the end. | vLLM in Docker |
Language models that reason step by step, call tools, and write code.
| Demo | What it does | Runs on |
|---|---|---|
| 🛠️ Agent Studio — Tool-Calling | Ask in plain language; the model decides which tools to call — calculator, clock, Python sandbox, knowledge search, vision — and reports back. | vLLM in Docker |
| 🧠 Reasoning AI — DeepSeek-R1 | Watch a reasoning model work: its private chain-of-thought streams into one panel while the final answer builds in another. | vLLM in Docker |
| 💻 AI Code Assistant | A private, self-hosted coding assistant that generates, explains, debugs and reviews code. No cloud, no code leaving the machine. | vLLM in Docker |
Read PDFs, scans, receipts, tables and forms; ask questions and get cited answers.
| Demo | What it does | Runs on |
|---|---|---|
| 📄 RAG Document Chatbot | Ask questions about your own PDFs. The notebook chunks and embeds the document, retrieves the passages that matter, and answers with citations. | vLLM in Docker |
| 📋 OCR & Document Intelligence | Turn scanned pages, receipts, tables, forms and handwriting into structured text, and ask questions about what the page says. | vLLM in Docker |
Turn a prompt into images, video, music and speech — the generative demos.
| Demo | What it does | Runs on |
|---|---|---|
| 📸 AI Photo Booth | Turn an idea or a photo into artwork in ten styles — a vision model, a language model and an image model working together on one Intel GPU. | vLLM in Docker |
| 🖌️ AI Art Studio | One sentence becomes a finished creative concept — a language model writes the brief, an image model paints it, a video model animates it. | PyTorch on-device |
| 🎙️ Voice AI Studio | A complete voice-to-voice assistant: speech in, reasoning in the middle, speech out — every stage on your Intel GPU. | vLLM in Docker |
Vision-language models that inspect an image and describe, report, diagnose or grade it.
| Demo | What it does | Runs on |
|---|---|---|
| 🔭 YOLO26 Vision Lab | One computer-vision model, five tasks — detect, segment, pose, oriented boxes and classify — on a single Intel Arc Pro card. | PyTorch on-device |
| 🏥 The Imaging Center — Radiology Assistant | A vision-language model reads a medical image and writes a structured report — findings, impression, and follow-up questions you can ask. | vLLM in Docker |
| 🔧 The Inspection Line — Defect Detection | A vision-language model inspects a manufactured part and returns a structured QC report with a PASS / FAIL / REWORK disposition. | vLLM in Docker |
| 🛒 The Smart Shelf — Product Recognition | Point the model at a product photo and get an identification, feature list, a price estimate and cross-sell recommendations. | PyTorch on-device |
Every demo has its own page with a pipeline diagram, its inputs and outputs, and the models it can run. The full reference table is in docs/CATALOG.md.
All demos follow the same shape, so once you have run one you can read any of them:
| Self-contained | The notebook checks the environment and installs missing Python packages itself. A working PyTorch installation is never changed. |
| Zero-click first result | Run All runs the pipeline once on a built-in sample, so you see real output before touching anything. |
| Interactive | Every demo has controls inline in the notebook. Most also open a full-screen app you can share with a colleague. |
| Bring your own input | Every demo accepts your own file, image, audio or text alongside the samples. |
| Yours to keep running | Models are served locally on your own GPU. Nothing is sent to a cloud service, and a cleanup cell releases the GPU when you are done. |
| docs/SETUP.md | Driver, Docker and Python setup, start to finish |
| docs/CATALOG.md | Every demo, model and container image in one table |
| docs/TROUBLESHOOTING.md | The problems people actually hit, and their fixes |
| docs/CONTRIBUTING.md | How to add a demo, and how releases work |
| CHANGELOG.md | What changed in each release |
| NOTICE | Third-party models, datasets and their licences |
| DISCLAIMER.md | Terms of use — evaluation only, provided "as is" |
| RESPONSIBLE_AI.md | Responsible use — GenAI & agentic best practices, human-rights notice |
These demos are provided for evaluation, validation, demonstration and educational purposes only. They are sample code intended to show how AI models can run on Intel® Arc™ Pro GPUs. They are not production-ready software, not a supported product or offering, and not intended for use in production, clinical, safety-critical, or other decision-making systems.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. TO THE MAXIMUM EXTENT PERMITTED BY LAW, THE AUTHORS AND COPYRIGHT HOLDERS DISCLAIM ALL LIABILITY FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY ARISING FROM, OUT OF, OR IN CONNECTION WITH THE SOFTWARE OR ITS USE.
By running these notebooks, you accept full responsibility for:
- Your environment and data — the notebooks run on your own machine and process the inputs you give them. You are responsible for securing that machine and for any data you choose to send to a model.
- Network exposure — some notebooks can open a public share link (for example a
*.gradio.liveURL) so the app is reachable from a browser. If you enable it, anyone with the link can reach the running app and any input you send it. Leave it off to stay on your local network. - Model outputs — AI models can produce inaccurate, incomplete or misleading results. No output is a substitute for professional judgement, and nothing here should be relied upon without independent verification.
- Third-party models and data — each model and dataset is downloaded from its upstream source under its own licence. Some require you to accept terms first (for example, gated models on Hugging Face). Complying with those licences, and with all applicable laws, is your responsibility. See NOTICE.
Nothing is collected from your use of these notebooks; everything runs locally on your hardware.
These demos generate content and include agentic and interactive examples. Please review AI outputs before relying on them, verify facts and references, and remember you are responsible for any action taken on generated content. Full best-practices for generative and agentic use are in RESPONSIBLE_AI.md.
Intel is committed to respecting human rights and avoiding complicity in human rights abuses. See Intel's Global Human Rights Principles. Intel's products and software are intended only to be used in applications that do not cause or contribute to a violation of an internationally recognized human right. Report any potentially inappropriate use to responsible-ai@intel.com.
Built by the Intel XPU AI demos team:
@akashdhamasia12 · @YuningQiu · @ctao456 · @yogit2020 · @GopeshKh · @gilliean · @ahmedintel · @mins2022 · @jooncho1 · @258SusanLiu · @vvravi1 · @dunniie · @zhiqitao-intel · @rdower · @RamaKaramsetty
The notebooks and tooling in this repository are licensed under Apache-2.0. The models and sample datasets they download are covered by their own licences — see NOTICE and each demo's page.