vLLM on Intel Arc Pro B-Series: Serving LLMs to a Whole Team (Official Path)

Ollama is the easy path for one person on one machine (our guide). But when a team hits the same model, or you are serving an app, you want vLLM: continuous batching, OpenAI-compatible API, real concurrency. As of late 2025 vLLM officially supports Intel Arc Pro B-Series, and Intel maintains a dedicated container stack for it. Here is the working path, every command from the official sources linked.

What you need

• An Arc Pro B-Series GPU: B70 32GB or B60 Dual 48GB (the Dual needs a motherboard doing PCIe x8/x8 bifurcation: read this first)
• Fresh Ubuntu 24.04 LTS (.3 or .4), kernel 6.8+, Docker
• Intel GPU driver installed and both dies visible (see the verification steps in the Ollama guide)

1. Get Intel's vLLM container

Intel's canonical B-series stack is llm-scaler-vllm (their Project Battlematrix software line). Important, from Intel directly: do not pull :latest; pick the exact current version from the Releases page.

docker pull intel/llm-scaler-vllm:<VERSION>
sudo docker run -td --privileged --net=host --device=/dev/dri \
  --name=lsv-container -v /home/<you>/models:/llm/models/ \
  --shm-size="32g" --entrypoint /bin/bash intel/llm-scaler-vllm:<VERSION>
docker exec -it lsv-container bash

Add yourself to the GPU groups first if you have not (sudo usermod -aG render $USER && sudo usermod -aG video $USER, then log out and in). Source: intel/llm-scaler vLLM README.

2. Serve a model

Inside the container. The flag that matters on the B60 Dual: --tensor-parallel-size 2 spans both GPU dies.

vllm serve --model /llm/models/<MODEL> --tensor-parallel-size 2 \
  --quantization fp8 --dtype=float16 --enforce-eager \
  --port 8000 --host 0.0.0.0 --max-model-len=8192

That exposes an OpenAI-compatible endpoint on port 8000: anything that speaks the OpenAI API (apps, agents, LangChain, your internal tools) points at it with a base URL change.

3. Benchmark it properly

Verbatim from Intel's README; reports time-to-first-token, per-token latency and total throughput:

vllm bench serve --model /llm/models/<MODEL> --dataset-name random \
  --served-model-name <MODEL> --random-input-len=1024 --random-output-len=512 \
  --ignore-eos --num-prompt 10 --trust_remote_code --request-rate inf \
  --backend vllm --port=8000

For scale context: the official vLLM blog measured a 4x B60 rig serving gpt-oss-120b at roughly 1,495 output tokens/sec across 100 concurrent requests, with near-linear multi-GPU scaling. A single Dual card is a quarter of that machine at a fraction of the price.

Honest notes

• This stack is Linux + Docker territory. If that is not your happy place, that is literally why we build these machines configured and benchmarked before they ship.
• Intel marks parts of the platform tooling as demo-grade; the vLLM serving itself is the officially supported path and moves quickly, so check the Releases page each install.
• Sizing: which models fit in 32GB vs 48GB is in the VRAM guide; the cost case vs cloud is in the payback calculator.

Deploying for a team or a client and want it to just arrive working? Email alex@destellotech.com with your models and user count. UK stock, VAT invoice, direct founder support.