Intel Arc for Local AI (2026): B580, B70, Ollama, ComfyUI

With NVIDIA cards selling far above MSRP, a lot of people are asking whether an Intel Arc GPU is good enough for local AI. The short answer is yes for LLMs, if you set it up the right way, and “with some effort” for ComfyUI.

This guide covers every current Arc card, from the 10 GB B570 to the 32 GB Arc Pro B70. It also shows the setup that works in late 2026 (Vulkan, not the old IPEX-LLM), real speeds, and the problems you’ll hit.

TL;DR

Best route for LLMs: llama.cpp’s Vulkan backend, which Ollama and LM Studio use on Arc. Skip IPEX-LLM: Intel archived it.
Arc B580 (12 GB, $249 MSRP): runs 7B-14B models. One user measured about 28 tok/s on Llama 3.1 8B Q4.
Arc Pro B60 (24 GB) and B70 (32 GB, from $949): the interesting cards for 27B-32B models.
ComfyUI: works with PyTorch XPU builds, but Intel support is still marked prototype and many custom nodes assume CUDA.
Verdict: great value for chat LLMs on a budget. For image and video work, NVIDIA is still much easier.

Intel Arc Cards for AI: Specs and Prices

For LLMs, VRAM decides what fits and bandwidth decides speed. Here’s every current Arc card that matters, from Intel’s spec pages.

Card	VRAM	Bandwidth	Power	Price
Arc B570	10 GB	380 GB/s	150 W	$219 MSRP, about $289 now
Arc B580	12 GB	456 GB/s	190 W	$249 MSRP, often out of stock
Arc Pro B50	16 GB	224 GB/s	70 W	$299 suggested
Arc Pro B60	24 GB	456 GB/s	200 W	No Intel price; sold via partners
Arc Pro B65	32 GB	608 GB/s	200 W	No Intel price published
Arc Pro B70	32 GB	608 GB/s	230 W	From $949 (Intel card)
Arc A770 16GB (older)	16 GB	560 GB/s	225 W	$349 MSRP, about $399 now

Street prices are from Tom’s Hardware’s GPU price tracker (late October 2026). Pro card details are from Aliteq’s summary of Intel’s pages.

Note: the Pro B50 has 16 GB but only 224 GB/s. It fits bigger models than a B580 but writes tokens at about half the speed.

What Models Fit on Each Card

These are 4-bit quants with a 16K context, based on Aliteq’s memory maths from published model configs. They’re fits, not speed tests.

Card	Fits comfortably	Tight or no
8 GB (A750, A580)	Qwen3.5 9B, 7B-8B models	Anything 12B+
B570 (10 GB)	Gemma 4 12B	14B models
B580 (12 GB)	Gemma 4 12B, 7B-9B at Q6-Q8	Qwen3 14B is tight; gpt-oss 20B doesn’t fully fit
Pro B50 / A770 (16 GB)	gpt-oss 20B, 14B with long context	27B+
Pro B60 (24 GB)	Gemma 4 31B	Qwen3 32B is tight
Pro B65 / B70 (32 GB)	Qwen3 32B with room for context	70B needs offload

To check any model against your card, use my VRAM requirements guide. For ready-made picks, see the best Ollama models by VRAM.

Vulkan vs SYCL vs IPEX-LLM

This is where most Arc guides go wrong. A lot of older tutorials still send you to IPEX-LLM, which is now a dead end.

Backend	Status in late 2026	Use it?
Vulkan (llama.cpp)	Maintained, used by Ollama and LM Studio on Arc	Yes, start here
SYCL (llama.cpp + oneAPI)	Maintained, verified on A770, A750 and B580	Optional, for advanced users
IPEX-LLM	Archived by Intel; README says the project is archived	No
Intel Extension for PyTorch (IPEX)	Archived, with known security warnings	No, use native PyTorch XPU

Warning: if a guide asks you to install ipex-llm or intel-extension-for-pytorch, it’s out of date. Use Vulkan for LLMs and native PyTorch XPU for ComfyUI.

llama.cpp’s September 2026 notes mention a newer oneAPI release that made SYCL prompt processing 3.1x faster on a B570. So SYCL is improving, but Vulkan is still the easiest path.

How Fast Is Intel Arc?

Measured Arc numbers are rare. The best I found is one user’s Arc B580 test with llama.cpp on Vulkan (published June 2026 on DEV).

Model	Quant	Speed	VRAM
Mistral 7B v0.3	Q4_K_M	31.4 tok/s	6.6 GB
Llama 3.1 8B Instruct	Q4_K_M	28.1 tok/s	6.5 GB
Gemma 2 9B	Q4_K_M	26.5 tok/s	7.4 GB
Llama 3.1 8B Instruct	Q5_K_M	23.8 tok/s	7.8 GB
13B-class model	Q4_K_M	17.2 tok/s	10.5 GB

Prompt processing on the 8B models was about 590-640 tok/s. These are self-reported results, so treat them as a rough guide rather than a lab benchmark.

Note: the B580 is slower than its bandwidth suggests. An NVIDIA card with similar bandwidth, like the RTX 5060 Ti, gets well over 50 tok/s on 7B models. Arc drivers and kernels are still catching up.

Bigger Pro cards scale with bandwidth. The B70’s 608 GB/s should make it roughly a third faster than a B580 on models that fit both, but I haven’t found measured B70 numbers yet.

Setup: Ollama and LM Studio on Arc

This is the easiest way to start. Both apps use llama.cpp’s Vulkan backend on Intel GPUs.

Update the Intel graphics driver. Get the latest Arc driver from Intel’s site (Windows) or your distro’s latest Mesa and compute runtime (Linux).
Check Vulkan sees the card. On Linux run vulkaninfo | grep deviceName. On Windows, the Arc control panel should show the current driver.
Install Ollama. Use the normal installer from my Ollama install guide.
Turn on Vulkan if needed. If Ollama runs on the CPU, set the environment variable OLLAMA_VULKAN=1 and restart Ollama.
Run a model. Try ollama run qwen3:8b, then check ollama ps. You want it to show the GPU, not 100% CPU.
Or use LM Studio. Install it, open the runtime settings and pick the Vulkan llama.cpp runtime. Then download a Q4_K_M model.

Tip: on Linux, watch the Ollama log with journalctl -f -u ollama. A working setup shows your Arc card as a Vulkan device with its memory size.

Not sure which app to use? My Ollama vs LM Studio vs llama.cpp comparison explains the trade-offs.

Setup: llama.cpp with Vulkan

For the most control, run llama.cpp directly. You can download a Vulkan build from the llama.cpp releases page or build it yourself.

Install Vulkan tools (Linux). On Ubuntu: sudo apt install vulkan-tools libvulkan-dev.
Build llama.cpp. Clone the repo, then run cmake -B build -DGGML_VULKAN=1 and cmake --build build --config Release.
Download a GGUF. For example a Q4_K_M quant of an 8B-14B model from Hugging Face.
Run it on the GPU. Use -ngl 99 to put all layers on the card. Newer builds use llama serve and llama cli.
Check speed. If you get under 10 tok/s on an 8B model, the layers are probably on the CPU.

ComfyUI on Intel Arc

ComfyUI officially supports Intel GPUs through native PyTorch XPU builds. It works, but expect more rough edges than on NVIDIA.

Install Python 3.13 and create a virtual environment.
Install PyTorch for XPU: pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/xpu
Clone ComfyUI and run pip install -r requirements.txt inside the ComfyUI folder.
Check the GPU. In Python, torch.xpu.is_available() should return True.
Start it with python main.py and open http://127.0.0.1:8188.

Out of memory? Add --reserve-vram 2 or --lowvram.
Don’t use --directml; ComfyUI itself warns it barely works.
FP8 models save memory on Arc, but the maths still runs in 16-bit, so there’s no FP8 speed-up.
Custom nodes that call CUDA directly won’t work. GGUF loaders may or may not work on Arc, so test first.

A 12 GB B580 can run SDXL, SD 3.5 Medium and a 4-bit GGUF of FLUX.1 dev by file size. For video models, an NVIDIA card is still the safer choice. My must-have ComfyUI custom nodes list shows how to install nodes.

Tip: if you don’t want a terminal, Intel’s AI Playground app runs image models and chat on Arc. It’s beta software and needs 8 GB+ of VRAM.

Intel Arc vs NVIDIA vs AMD for AI

Card	VRAM	Price (late Oct 2026)	Strength	Weakness
Intel Arc B580	12 GB	$249 MSRP (scarce)	Cheapest 12 GB card	Slower kernels, fewer tools
Intel Arc Pro B70	32 GB	From $949	One of the cheapest new 32 GB cards	Pro card, limited availability
RTX 5060 Ti 16GB	16 GB	~$765	CUDA, everything just works	Price far above MSRP
Radeon RX 9060 XT 16GB	16 GB	~$599	ROCm and Vulkan both work	Fewer ComfyUI nodes than CUDA

Quick answer: for chat LLMs on a tight budget, Arc is fine. A B70 is one of the cheapest ways to get 32 GB on one new card. If you mainly use ComfyUI, pay the extra for NVIDIA.

See my AMD GPU for AI guide, RTX 5060 Ti 16GB guide and best GPU for local LLMs for the other options.

Troubleshooting

Ollama uses the CPU, not the Arc card

Update the driver, set OLLAMA_VULKAN=1, and restart Ollama. Check the log for a Vulkan device line, then run ollama ps.

Very slow generation

Under 10 tok/s on an 8B model usually means CPU fallback. In llama.cpp, add -ngl 99. Also check the model plus context fits in VRAM.

torch.xpu.is_available() returns False

The driver or compute runtime isn’t installed properly. Reinstall the latest Arc driver, then make sure you installed the XPU PyTorch build, not the CPU one.

A guide tells you to install IPEX-LLM

It’s out of date. IPEX-LLM is archived, so use Vulkan llama.cpp or native PyTorch XPU instead.

Crashes or flicker on Linux

One Fedora user stopped crashes on a B580 by removing the intel-media-driver and intel-gmmlib packages. They fixed Flatpak app flicker with GSK_RENDERER=cairo.

Frequently Asked Questions

Is Intel Arc good for AI?

Yes for local LLMs through Vulkan in Ollama, LM Studio or llama.cpp. For ComfyUI it works but is harder, because many tools assume CUDA.

Is the Intel Arc B580 good for local LLMs?

For 7B-14B models, yes. One user measured about 28 tokens per second on Llama 3.1 8B at Q4 using llama.cpp on Vulkan.

Does Ollama support Intel Arc GPUs?

Yes, through its Vulkan backend. If it runs on the CPU, set OLLAMA_VULKAN=1 and restart Ollama.

Should I use IPEX-LLM?

No. Intel archived IPEX-LLM, and its README says so. Use llama.cpp with Vulkan or SYCL instead.

Can I run ComfyUI on Intel Arc?

Yes. Install the PyTorch XPU build from download.pytorch.org/whl/xpu, then ComfyUI as normal. Some custom nodes won’t work.

Which Intel Arc card is best for AI?

The Arc Pro B70 with 32 GB, from $949, for big models. The B580 at $249 MSRP is the budget pick for 7B-14B models.

Intel Arc B580 or RTX 5060 Ti for AI?

The 5060 Ti has 16 GB, CUDA and faster speeds, but costs about three times as much. Arc is the budget choice for chat LLMs.

Final Thoughts

Intel Arc is a real option for local AI in 2026, as long as you use the right software. Vulkan in Ollama, LM Studio or llama.cpp is the path that works today.

The B580 is a cheap way to run 7B-14B models, and the Pro B70 is one of the cheapest new 32 GB cards. If ComfyUI is your main use, stick with NVIDIA, or rent a GPU for heavy jobs with my cheapest cloud GPU guide.

Intel Arc GPU for Local AI (2026): B580, B60, B70 with Ollama, LM Studio, llama.cpp and ComfyUI