Intel's Arc graphics cards are best known as good-value gaming GPUs, but they can also run large language models (LLMs) and image generators on your own PC, with no subscription and no data leaving your machine. Most local AI guides assume an NVIDIA card, though, so Arc owners often have to piece together the steps.
This guide is based on official documentation from Intel, Ollama, LM Studio and Jan, checked on 14 November 2026. Forge PC has not benchmarked these apps on Arc hardware, so we do not quote speeds; we explain which tools support Arc and how to set them up.
What you need
An Intel Arc graphics card. Intel's AI Playground needs an Arc A-series or B-series discrete card with at least 8GB of video memory for its full feature set. The popular Arc B580 has 12GB of GDDR6 memory, which Intel quotes at 456GB/s, plus 160 XMX matrix engines for AI work.
Windows 11 (or Windows 10 for some apps), with a recent Intel graphics driver.
16GB of system RAM is the sensible minimum. LM Studio recommends at least 16GB, and Jan lists 8GB as the minimum and 16GB as recommended. Our RAM for local AI guide explains why more helps.
Plenty of storage. Models are large files; Intel recommends 82GB or more of free space for AI Playground and its models.
Step 1: Update your Intel graphics driver
Every tool below depends on a current driver. Download the latest Intel Arc Graphics driver for Windows from Intel's website; on 14 November 2026 Intel listed version 32.0.101.9175, which supports Arc B-series and A-series cards on Windows 10 (22H2) and Windows 11. Our graphics driver update guide covers the process. To check the card is recognised, open Device Manager and look under Display adapters; Intel's AI Playground instructions say it should list an Intel Arc device.
Step 2: Pick an app
App | Best for | How it uses Arc (per official docs) |
|---|
Intel AI Playground | An all-in-one Intel app for chat and image generation | llama.cpp (Vulkan) and OpenVINO for chat; ComfyUI for images |
Ollama | A simple background service with a command line and API | Vulkan backend, enabled by default |
LM Studio | A polished desktop app for browsing and chatting with models | llama.cpp-based runtimes; check which runtimes it offers for your card |
Jan | An open-source desktop app | Bundled Vulkan backend for Intel Arc |
Intel AI Playground
AI Playground is Intel's own free, open-source app for chat, image generation and other AI features on Intel hardware. According to its GitHub page, it runs GGUF chat models through llama.cpp's Vulkan backend and also supports OpenVINO models. It installs ComfyUI for image and video generation.
Download the installer from the releases section of Intel's AI Playground GitHub page. Intel labels current versions as beta.
On first run, the setup wizard asks you to choose a hardware mode. Choose Studio for an Arc discrete card; the Essentials mode targets Intel Core Series 3 laptops and offers a smaller feature set.
Install the backend components the wizard offers (llama.cpp, OpenVINO and ComfyUI), then download a model from the built-in Model Manager.
Ollama
Ollama's documentation says that on Windows, support for other GPU vendors is provided through Vulkan, which "is enabled by default", and that most vendors' Windows drivers include Vulkan with no extra setup. After installing Ollama, run a model (for example ollama run followed by a model name), then type ollama ps in a second terminal window. The Processor column should say "100% GPU" if the model is fully loaded onto your card. Our Ollama and LM Studio guide covers installation in detail.
LM Studio
LM Studio runs on Windows PCs with an AVX2-capable processor and recommends at least 4GB of dedicated video memory. Its runtimes are based on llama.cpp. Once installed, open its runtime settings to see which engines are available for your card and keep them updated, then offload as many model layers to the GPU as your VRAM allows.
Jan
Jan's Windows documentation lists support for NVIDIA, AMD and Intel Arc GPUs, and says Intel Arc cards need Vulkan support. Jan bundles the Vulkan backend, so after installing a current Intel driver, go to Settings > Hardware > GPUs and make sure your Arc card is switched on. Jan lists 6GB of VRAM as its minimum for GPU acceleration.
What can 12GB of VRAM run?
Video memory decides which models you can run comfortably. As a rule of thumb, the model file and its conversation memory (the KV cache) need to fit in VRAM for full speed. Quantised GGUF models make that much easier: our quantisation explainer and VRAM guide show how to estimate sizes. On a 12GB card such as the B580, small and mid-sized models in compact quantisations are the comfortable range; larger models will spill into system RAM and slow down.
If you want more memory for AI, Intel also sells the Arc Pro B70, which PC Gamer reported in March 2026 has 32GB of GDDR6 and a $949 US price. It is aimed at AI and workstation use rather than gaming.
Is an Arc B580 a good AI buy?
SpecPrice's UK tracker listed the Arc B580 from £305.70 on 14 November 2026 (approximate; prices change daily). That makes it one of the cheapest new 12GB cards. Software support is the trade-off: most local AI tools are developed first for NVIDIA's CUDA, and Arc support arrives through Vulkan, SYCL and OpenVINO. The llama.cpp 1.0 release notes, published on 13 November 2026, include Intel-specific Vulkan work, a sign that support is improving. Compare options in our best graphics cards for local AI and best budget graphics card guides, and if you have an AMD card, see our Radeon local AI guide.
Our verdict
You can run local AI on an Intel Arc GPU today: update the Intel graphics driver, then use Intel's AI Playground in Studio mode for an all-in-one app, or Ollama, LM Studio or Jan, which use Vulkan to reach Arc cards. A 12GB Arc B580, listed from about £305.70 in the UK on 14 November 2026, suits small and mid-sized quantised models, but expect software support to trail NVIDIA's. Forge PC has not benchmarked these apps on Arc; this guide follows Intel's and each app's official documentation.