How to Run Ollama on an Intel Arc GPU (B60, B70): The Official Path

Yes, Ollama runs properly on Intel Arc. Intel maintains an official Ollama build for Arc GPUs through its IPEX-LLM project, verified on the Arc B-Series (that includes the Arc Pro B60 24GB, B70 32GB and B60 Dual 48GB). Here is the whole path on Ubuntu, every command from Intel's own documentation, linked as we go.

1. OS and kernel

Use Ubuntu 24.04 LTS (which tracks the HWE kernel; Battlemage cards need kernel 6.8 or newer) or Ubuntu 25.10+. Check with uname -r.

2. Install the GPU driver

From Intel's official client GPU guide (dgpu-docs):

sudo apt-get update
sudo apt-get install -y software-properties-common
sudo add-apt-repository -y ppa:kobuk-team/intel-graphics
sudo apt-get install -y libze-intel-gpu1 libze1 intel-metrics-discovery intel-opencl-icd clinfo intel-gsc
sudo gpasswd -a ${USER} render
newgrp render

Verify: clinfo | grep "Device Name" should list your Arc GPU. On a B60 Dual you should see it twice, once per die; also check lspci -nn | grep -Ei 'VGA|DISPLAY' shows two Battlemage entries. Only one? Read our bifurcation guide.

3. Get Intel's Ollama build

Download the latest Linux portable tgz from the IPEX-LLM Ollama quickstart (release asset named like ollama-ipex-llm-*-ubuntu.tgz). No compiling, no CUDA, no drama:

tar -xvf ollama-ipex-llm-*-ubuntu.tgz
cd <extracted folder>
export OLLAMA_NUM_CTX=16384
./start-ollama.sh

The default context is only 2048 tokens, hence the export. On a multi-GPU or dual-die card, pin the devices before starting: export ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"

4. Run a model and get your tokens per second

# second terminal, same folder:
./ollama run deepseek-r1:7b --verbose

--verbose prints eval rate: N tokens/s after each response. That is your benchmark number. Which model sizes fit which card: see the VRAM guide.

Known limits, honestly

Intel's portable build tracks Ollama a few versions behind upstream, and for serious multi-user serving you want vLLM instead (Intel maintains an official Arc B-series vLLM stack, which is what we use for benchmarking). For a single user running local models, the portable Ollama build is the easiest working path on Arc today.

Want the hardware without the setup? We stock the Arc Pro range in the UK and build complete machines with all of the above preinstalled and benchmarked before dispatch. Email alex@destellotech.com with any setup question, buyer or not, and you will get a straight answer.