What is the best Intel GPU for local LLMs?

For the most models, the Intel Arc Pro B70 32GB: 32 GB of VRAM and 608 GB/s for $949 list, enough for 27B–32B-class models at Q4 (Qwen3.8-27B at an estimated 22–31 tok/s). The cheapest current GeForce with 32 GB is the GeForce RTX 5090 at $1,999 list. On a budget, the Intel Arc B580 (12 GB, $249 list) runs 8B–14B models.

Do Intel Arc GPUs work for running LLMs?

Yes. llama.cpp runs on Arc through its Vulkan backend (what LM Studio and Ollama use) and its SYCL backend (Intel's oneAPI path, which lists the Arc A770, A750 and B580 as verified devices). Intel's OpenVINO GenAI and vLLM's Intel GPU (XPU) backend are other options. Expect fewer tools and tutorials than for NVIDIA, and more sensitivity to driver versions.

Should I use IPEX-LLM?

Not for new setups. Intel archived the IPEX-LLM repository on January 28, 2026, so it no longer gets updates. Many 2024–2025 guides still point to it. Use llama.cpp (Vulkan or SYCL), LM Studio, Ollama or OpenVINO instead.

Is the Arc B580 good for local LLMs?

For the price, yes: 12 GB and 456 GB/s for $249 list, enough for 8B models with room for context (Qwen3-8B at an estimated 55–78 tok/s at Q4). 14B models fit at Q4 too, but only with a short context. The cheapest 12 GB GeForce we track is the GeForce RTX 3060 12GB at $329 list.

Is a used Arc A770 16GB worth it?

It has 16 GB and 560 GB/s, more than the B580, so it fits 14B models with more context. It's discontinued (launch MSRP was $349), so you'll be buying used. The older Alchemist chip idles at higher power and gets less tuning in llama.cpp than Battlemage. New, the cheapest 16 GB GeForce we track is the GeForce RTX 5060 Ti 16GB at $429 list.

Intel, AMD or NVIDIA for local AI?

Intel and AMD give more VRAM per dollar; NVIDIA has the widest software support (CUDA, FlashAttention, most fine-tuning tools). For chat and coding in llama.cpp-based apps, all three work. Intel's 32 GB Arc Pro B70 is one of the cheapest ways to get 32 GB on a single new card. See NVIDIA vs AMD vs Mac and the best AMD GPU guide.

Best Intel GPU for local LLMs

Intel Arc cards are some of the cheapest ways to get 12 to 32 GB of VRAM. Here are the 5 Arc GPUs we track, with list prices, estimated tokens per second, which software runs them, and the cheapest NVIDIA and AMD card with the same VRAM.

Intel GPUs tracked: 5
Still sold new: 3
VRAM range: 8–32 GB

Short answer: the Arc Pro B70 (32 GB, $949 list) is the best Intel card for local LLMs and fits 27B–32B models at Q4. The Arc B580 (12 GB, $249) is the budget pick for 8B–14B models. A used A770 16GB is the cheapest way into 16 GB on Intel.

About software: llama.cpp supports Arc through Vulkan and SYCL (SYCL backend docs); LM Studio and Ollama use Vulkan. Intel's old IPEX-LLM project was archived on 2026-01-28 (checked 2026-11-10), so skip guides that rely on it. Prices are list prices (MSRP) or launch MSRPs for discontinued cards, never street prices.

Our picks

Most VRAM: 32 GB for 27B–32B models

Intel Arc Pro B70 32GB

VRAM: 32 GB
Bandwidth: 608 GB/s
Run at Q4: 504 / 656

Battlemage (Xe2). Qwen3-32B at Q4: est. 18–26 tok/s. Biggest at Q4: IQuest-Coder-V1-40B-Loop-Instruct.

$949 (list price (MSRP), source)
Cheapest GeForce with 32+ GB: GeForce RTX 5090, $1,999 list.

Amazon eBay

Best budget card: 12 GB

Intel Arc B580

VRAM: 12 GB
Bandwidth: 456 GB/s
Run at Q4: 350 / 656

Battlemage (Xe2). Qwen3-8B at Q4: est. 55–78 tok/s. Biggest at Q4: OLMo-2-1124-13B-Instruct.

$249 (list price (MSRP), source)
Cheapest GeForce with 12+ GB: GeForce RTX 3060 12GB, $329 list.

Amazon eBay

16 GB on the used market

Intel Arc A770 16GB

VRAM: 16 GB
Bandwidth: 560 GB/s
Run at Q4: 388 / 656

Alchemist (Xe). Qwen3-14B at Q4: est. 38–53 tok/s. Biggest at Q4: Ling-Coder-lite.

$349 (launch MSRP, discontinued, source)

Amazon eBay

Cheapest current Arc: 10 GB

Intel Arc B570

VRAM: 10 GB
Bandwidth: 380 GB/s
Run at Q4: 340 / 656

Battlemage (Xe2). Qwen3-8B at Q4: est. 46–65 tok/s. Biggest at Q4: HyperCLOVAX-SEED-Omni-8B.

$219 (list price (MSRP), source)
Cheapest GeForce with 10+ GB: GeForce RTX 3060 12GB, $329 list.

Amazon eBay

We may earn a commission when you buy through these links, at no extra cost to you.

Every Intel GPU we track

Sorted by memory, then bandwidth. Board power is the reference spec (how we calculate fit).

Intel Arc GPUs ranked by VRAM with bandwidth, board power, price, price per GB and the biggest model each runs at Q4
GPU	VRAM	Bandwidth	Board power	Price	$ per GB	Biggest model at Q4	Models that fit
Intel Arc Pro B70 32GBBattlemage (Xe2)	32 GB	608 GB/s	230 W	$949list price (MSRP), source, 2026-10-29	$30	IQuest-Coder-V1-40B-Loop-Instruct~28.7 GB at Q4	504 / 656
Intel Arc A770 16GBAlchemist (Xe) · discontinued, buy used	16 GB	560 GB/s	225 W	$349launch MSRP, discontinued, source, 2026-10-28	$22	Ling-Coder-lite~12.7 GB at Q4	388 / 656
Intel Arc B580Battlemage (Xe2)	12 GB	456 GB/s	—	$249list price (MSRP), source, 2026-10-22	$21	OLMo-2-1124-13B-Instruct~10.6 GB at Q4	350 / 656
Intel Arc B570Battlemage (Xe2)	10 GB	380 GB/s	150 W	$219list price (MSRP), source, 2026-10-28	$22	HyperCLOVAX-SEED-Omni-8B~8.5 GB at Q4	340 / 656
Intel Arc A750Alchemist (Xe) · discontinued, buy used	8 GB	512 GB/s	225 W	$289launch MSRP, discontinued, source, 2026-10-28	$36	NVIDIA-Nemotron-Nano-9B-v2~7.2 GB at Q4	315 / 656

Which popular models run, and how fast

Estimated decode speed at Q4 in tokens per second (how we estimate speed); MoE models show an upper bound only. Real Arc speeds depend on the backend (Vulkan or SYCL) and driver version.

Estimated fit and decode speed at Q4 for popular models on each Intel Arc GPU
Intel GPU	Llama-3.1-8B-Instruct ~6.6 GB	Qwen3-8B ~6.7 GB	Qwen3-14B ~11.3 GB	gpt-oss-20b ~15.6 GB	Qwen3-30B-A3B ~22.3 GB	Qwen3.8-27B ~20.4 GB	Qwen3-32B ~23.8 GB
Intel Arc Pro B70 32GB32 GB · 608 GB/s	Fits 75–106 tok/s	Fits 73–104 tok/s	Fits 41–58 tok/s	Fits up to 237 tok/s	Fits up to 284 tok/s	Fits 22–31 tok/s	Fits 18–26 tok/s
Intel Arc A770 16GB16 GB · 560 GB/s	Fits 69–98 tok/s	Fits 68–96 tok/s	Fits 38–53 tok/s	Tight up to 218 tok/s	Won’t fit	Won’t fit	Won’t fit
Intel Arc B58012 GB · 456 GB/s	Fits 56–80 tok/s	Fits 55–78 tok/s	Tight 31–43 tok/s	Won’t fit	Won’t fit	Won’t fit	Won’t fit
Intel Arc B57010 GB · 380 GB/s	Fits 47–66 tok/s	Fits 46–65 tok/s	Won’t fit	Won’t fit	Won’t fit	Won’t fit	Won’t fit
Intel Arc A7508 GB · 512 GB/s	Fits 63–89 tok/s	Fits 62–88 tok/s	Won’t fit	Won’t fit	Won’t fit	Won’t fit	Won’t fit

Intel vs NVIDIA vs AMD at the same VRAM

For each memory size Intel still sells: the cheapest current Arc, GeForce and Radeon card we track with at least that much VRAM, all at list price (MSRP).

Cheapest current Intel, NVIDIA and AMD GPU at each VRAM size, at list price
VRAM	Cheapest Intel	Cheapest NVIDIA	Cheapest AMD
10+ GB	Intel Arc B57010 GB · 380 GB/s · $219 list	GeForce RTX 3060 12GB12 GB · 360 GB/s · $329 list	Radeon RX 7600 XT16 GB · 288 GB/s · $329 list
12+ GB	Intel Arc B58012 GB · 456 GB/s · $249 list	GeForce RTX 3060 12GB12 GB · 360 GB/s · $329 list	Radeon RX 7600 XT16 GB · 288 GB/s · $329 list
32+ GB	Intel Arc Pro B70 32GB32 GB · 608 GB/s · $949 list	GeForce RTX 509032 GB · 1792 GB/s · $1,999 list	None we track

Before you buy Intel

Start with Vulkan. LM Studio and Ollama work with just the GPU driver. Try llama.cpp's SYCL build (oneAPI) if you need faster prompt processing on long inputs.
Keep drivers current. Arc performance in llama.cpp has improved a lot with newer drivers and builds; old benchmarks undersell the cards.
Enable Resizable BAR. Intel recommends it for Arc cards; check your motherboard BIOS.
Fine-tuning and CUDA-only tools mostly target NVIDIA. If you train models or use bleeding-edge inference kernels, see the best GPU guide.
Power and fit: check board power and case size in the planner.

Head-to-head: Arc B580 vs RTX 3060 12GB · Arc A770 vs RTX 4060 Ti 16GB · Arc Pro B70 vs RTX 3090 · Arc Pro B70 vs RX 7900 XTX · Arc Pro B70 vs RTX 5090.

Quick answers

What is the best Intel GPU for local LLMs?: For the most models, the Intel Arc Pro B70 32GB: 32 GB of VRAM and 608 GB/s for $949 list, enough for 27B–32B-class models at Q4 (Qwen3.8-27B at an estimated 22–31 tok/s). The cheapest current GeForce with 32 GB is the GeForce RTX 5090 at $1,999 list. On a budget, the Intel Arc B580 (12 GB, $249 list) runs 8B–14B models.
Do Intel Arc GPUs work for running LLMs?: Yes. llama.cpp runs on Arc through its Vulkan backend (what LM Studio and Ollama use) and its SYCL backend (Intel's oneAPI path, which lists the Arc A770, A750 and B580 as verified devices). Intel's OpenVINO GenAI and vLLM's Intel GPU (XPU) backend are other options. Expect fewer tools and tutorials than for NVIDIA, and more sensitivity to driver versions.
Should I use IPEX-LLM?: Not for new setups. Intel archived the IPEX-LLM repository on January 28, 2026, so it no longer gets updates. Many 2024–2025 guides still point to it. Use llama.cpp (Vulkan or SYCL), LM Studio, Ollama or OpenVINO instead.
Is the Arc B580 good for local LLMs?: For the price, yes: 12 GB and 456 GB/s for $249 list, enough for 8B models with room for context (Qwen3-8B at an estimated 55–78 tok/s at Q4). 14B models fit at Q4 too, but only with a short context. The cheapest 12 GB GeForce we track is the GeForce RTX 3060 12GB at $329 list.
Is a used Arc A770 16GB worth it?: It has 16 GB and 560 GB/s, more than the B580, so it fits 14B models with more context. It's discontinued (launch MSRP was $349), so you'll be buying used. The older Alchemist chip idles at higher power and gets less tuning in llama.cpp than Battlemage. New, the cheapest 16 GB GeForce we track is the GeForce RTX 5060 Ti 16GB at $429 list.
Intel, AMD or NVIDIA for local AI?: Intel and AMD give more VRAM per dollar; NVIDIA has the widest software support (CUDA, FlashAttention, most fine-tuning tools). For chat and coding in llama.cpp-based apps, all three work. Intel's 32 GB Arc Pro B70 is one of the cheapest ways to get 32 GB on a single new card. See NVIDIA vs AMD vs Mac and the best AMD GPU guide.

Buying guides · Best GPU (all brands) · Best AMD GPU · Best used GPU · Best GPU by budget · Tokens per second

Best Intel GPU for local LLMs

Our picks

Intel Arc Pro B70 32GB

Intel Arc B580

Intel Arc A770 16GB

Intel Arc B570

Every Intel GPU we track

Which popular models run, and how fast

Intel vs NVIDIA vs AMD at the same VRAM

Before you buy Intel

Quick answers

New models, and what they need to run