Ollama on Intel Arc B390: Local LLMs on the ASUS NUC 16 Pro

Running Ollama on Intel Arc graphics is now a mainstream path rather than a science project, and the ASUS NUC 16 Pro with the Core Ultra X9 388H or X7 358H is the smallest Windows machine we sell in Singapore that does it well: twelve Xe3 cores of Arc B390 graphics, LPDDR5X-9600 memory shared with the GPU, and a 50 TOPS NPU in a 0.7-litre box. This guide covers which NUC 16 Pro to buy, installing Ollama and LM Studio on Windows 11 with GPU acceleration, where Intel's acceleration stacks stand in 2026, which models fit 32 GB versus 64 GB, what the NPU is for, and when to step up to an Ascent GX10.

Which NUC 16 Pro for local LLMs

For LLM work the choice is the X-series. The Arc B390 has three times the Xe3 cores of the 356H's graphics (Intel rates the full Panther Lake GPU at up to 120 TOPS) and LPDDR5X-9600 feeds it faster than SODIMMs. The 356H's advantage is capacity: SODIMMs go to 128 GB, which loads a bigger file through a smaller GPU and slower memory, so it is the better office machine and the worse inference machine. Buy the X9 388H with 64 GB if models are the point; the 32 GB X7 358H is the value pick for 7B to 14B-class assistants. Our NUC buyer's guide covers the rest of the range.

Memory is the model ceiling, not TOPS

The Arc B390 has no VRAM of its own. It borrows system memory, so the 32 GB or 64 GB you buy is shared between Windows, your applications and the model weights plus context. Ollama's library gives the file sizes that matter: gpt-oss:20b is 14 GB, llama3.3:70b is 43 GB, gpt-oss:120b is 65 GB. On a 32 GB NUC 16 Pro, 7B to 14B-class models at 4-bit quantisation (roughly 4 to 9 GB) and the 14 GB gpt-oss:20b are comfortable with room for a long context; a 30B-class model at 4-bit is the practical ceiling. On 64 GB, 30B-class models become the daily driver, and a 43 GB 70B checkpoint will load, but with Windows and the context on top it is tight and slow; treat 70B as an experiment rather than a workflow. gpt-oss:120b at 65 GB does not fit either configuration. If 70B is the actual requirement, that is the GX10's job.

Installing Ollama on Windows 11 with Arc acceleration

Ollama's documented GPU backends are CUDA, ROCm, Metal and Vulkan, and Vulkan is the one that drives Intel graphics. It arrived as an experimental option in Ollama 0.12.6 in October 2025, and the current documentation states that Vulkan is enabled by default when the backend is installed, with no extra setup on Windows because most vendors' drivers already bundle Vulkan. The sequence on a NUC 16 Pro:

1. Install the latest Intel Arc graphics driver from Intel or ASUS support, then reboot. 2. Download OllamaSetup.exe from ollama.com; it needs Windows 10 22H2 or newer, does not require administrator rights and installs to your home directory. 3. Ollama runs in the background and serves its API on http://localhost:11434; open PowerShell and run ollama run gpt-oss:20b for a first chat. 4. Confirm the GPU is in use with ollama ps, which reports how much of the model sits on GPU versus CPU. 5. If you ever need to force CPU-only, set OLLAMA_VULKAN=0; to pin a specific Vulkan device, use GGML_VK_VISIBLE_DEVICES. 6. Models land in %HOMEPATH%\.ollama; on a NUC 16 Pro with a second M.2 drive fitted, set the OLLAMA_MODELS user environment variable to a folder on that drive, then quit and relaunch Ollama.

Any OpenAI-compatible client, VS Code extension or NVIDIA PAIR node can then point at port 11434.

LM Studio as the alternative

LM Studio bundles the llama.cpp engine behind a desktop interface with a model browser, and on Windows it ships a Vulkan llama.cpp runtime alongside a CPU runtime, selectable in its runtime settings. System requirements are a 64-bit CPU with AVX2 (any Core Ultra qualifies), 16 GB of RAM recommended and 4 GB of GPU memory recommended; it also supports Arm64 Windows, which matters if you are comparing against the Ascent QN10. Pick 4-bit GGUF builds, enable GPU offload and check the runtime is Vulkan rather than CPU if generation looks slow.

Where Intel's own acceleration stacks stand in 2026

Buyers who researched this a year ago will remember IPEX-LLM, Intel's library with "portable zip" builds of Ollama and llama.cpp for Arc. Intel archived that repository on 28 January 2026, read-only and unsupported; do not build a deployment on it. Three supported paths remain. Vulkan, through upstream Ollama, LM Studio or llama.cpp, is the default and needs nothing beyond the graphics driver. llama.cpp's SYCL backend targets Intel GPUs including Arc through oneAPI and suits people who compile their own builds. OpenVINO is Intel's own inference toolkit across CPU, GPU and NPU, and OpenVINO 2026.1 (8 April 2026) introduced a preview OpenVINO backend for llama.cpp "which enables optimized inference on Intel CPUs, GPUs, and NPUs"; it is labelled preview, validated so far on small GGUF models, and worth watching rather than depending on today.

What the 50 TOPS NPU does, and does not, do

Panther Lake's NPU is rated at 50 TOPS, comfortably above the 40 TOPS Copilot+ PC threshold, and Intel quotes about 180 platform TOPS when the GPU's 120 are added. Its job is the Windows side of AI: Copilot+ features and applications built on Windows ML or OpenVINO that compile a model for the NPU and run it at a few watts. Ollama and LM Studio do not use it; their backends are GPU and CPU only, so a chat model on the NUC 16 Pro runs on the Arc B390. That is a sensible division: the NPU handles always-on, low-power inference inside Windows, the GPU handles the 10 GB-plus language models, and both share one memory pool, which is one more argument for 64 GB.

Realistic expectations

A NUC 16 Pro will not match a discrete RTX card or a GB10 box. What you get is interactive speeds on 7B to 14B-class models, usable speeds on 20B to 30B-class, and a 70B that works but tests patience, with generation speed governed by memory bandwidth rather than the TOPS figure on the box. We do not publish tokens-per-second figures because they vary with quantisation, context length and driver version; measure your own models on your own prompts before committing a workflow. What the machine is excellent at is the single-user case: a coding assistant in VS Code, private document Q&A over a department's files, an agent that drafts and summarises, all on a Windows desktop under PDPA control with a three-year commercial warranty.

When to step up to the Ascent GX10

Move to the Ascent GX10 (128 GB unified memory, from S$7,200) when the model is 70B parameters or larger and must run daily, when more than one person will hit the endpoint at once, when you need CUDA for PyTorch training or fine-tuning, or when you want NVIDIA's NemoClaw sandboxed agents. Many of our customers run both: a NUC 16 Pro on each desk for Windows, Copilot+ and a small local assistant, and one GX10 as the shared 70B endpoint. The full ladder from NUC to DGX Station is in the ASUS Agent Computer guide.

Buying the NUC 16 Pro for local AI in Singapore

SourceIT is an authorised ASUS reseller with the NUC 16 Pro in X7 358H 32 GB / 1 TB at S$2,935, X9 388H 32 GB / 1 TB at S$3,390 and X9 388H 64 GB / 1 TB at S$4,420, all with Windows 11 Pro, official ASUS Singapore warranty and the three-year commercial warranty on business NUC models. We can deliver units with the Arc driver, Ollama and your chosen models pre-installed. Use Request for Quotation on the product page, call (65) 6978 3502, or email Sales@sourceit.com.sg; we are GeBIZ and Vendors@Gov registered.

Frequently asked questions

Does Ollama support Intel Arc graphics on Windows? Yes, through its Vulkan backend, which Ollama's documentation says is enabled by default on Windows once the Intel graphics driver is installed.

Do I need IPEX-LLM? No. Intel archived the IPEX-LLM repository on 28 January 2026. Use upstream Ollama, LM Studio or llama.cpp with Vulkan.

Can the NUC 16 Pro run a 70B model? A 4-bit 70B file is about 43 GB, so it loads on the 64 GB X9 388H but leaves little headroom and runs slowly. Treat 30B-class as the practical ceiling and use a GX10 for 70B.

Does Ollama use the 50 TOPS NPU? No. Ollama's backends are CUDA, ROCm, Metal and Vulkan. The NPU serves Copilot+ features and Windows ML or OpenVINO applications.

Is the 356H with 128 GB better for LLMs than the X9 with 64 GB? No. It holds a larger file but has 4 Xe3 cores and slower SODIMM memory. For inference the X9 388H's Arc B390 and LPDDR5X-9600 win.

Can other PCs use the NUC's models? Yes. Ollama serves an API on port 11434; open it on the LAN or add the NUC as a node in NVIDIA PAIR.

How fast is it? Interactive for 7B to 14B-class models, slower as size rises. Speed depends on quantisation, context and driver, so test your own models before committing.