Intel Arc B580 for Local AI: 12GB at $249 — Setup, Real Speeds, and Honest Limits

The Intel Arc B580 (12GB, $249 MSRP) is the best new budget GPU for local AI — published results show 60-80 tok/s on 7-8B models and 32-38 tok/s on 14B models under Ubuntu. But the software story flipped in 2026: Intel archived IPEX-LLM in January, and the recommended path today is Vulkan, which official Ollama now enables by default.

That second sentence is the part most B580 guides still get wrong, because they were written in 2025. This page covers what the card actually does (with sources for every number), the two setup paths that work in August 2026, why you should ignore the SYCL instructions you will find elsewhere, and the honest limits of a 12GB card. If you want the one-line purchase advice: at anything near $249, buy it; at heavily inflated street prices, a used 12GB card deserves a look first.

---

Why the B580 Matters {#why-b580}

Answer first: the B580 gives you 12GB of VRAM and 456 GB/s of bandwidth for a $249 MSRP — the cheapest new-card ticket to running 14B-class models entirely on GPU.

The spec sheet, per Intel's published specifications:

Why those numbers matter for local AI specifically: token generation speed is mostly a memory-bandwidth game, and 456 GB/s at this price is strong — the RTX 4060's 8GB/272 GB/s looks anemic next to it. And 12GB is the practical floor for running 14B models at Q4 without spilling into system RAM (a 14B Q4 needs roughly 9GB; our model RAM/VRAM table has the full breakdown by quant).

One honest caveat about price. The B580 sold out repeatedly through 2025, and the 2026 memory shortage has pushed street prices above MSRP across every GPU tier. $249 is Intel's list price, not a promise of what you'll pay this week — check live listings before you anchor on it. The buying advice in this guide assumes near-MSRP pricing; the further above $249 the card drifts, the more seriously you should compare a used 12GB RTX 3060.

---

The 2026 Software Story: IPEX-LLM Is Archived, Vulkan Won {#software-story}

Answer first: Intel archived the IPEX-LLM repository on January 28, 2026. The maintained path for Arc local AI is now Vulkan — official Ollama ships with the Vulkan backend enabled by default, and llama.cpp's community Vulkan numbers on the B580 now sit in the same band as Intel's own stack posted.

This is the section to read even if you skip everything else, because most B580 content online describes a software landscape that no longer exists.

The 2025 story was: vanilla Ollama did not support Arc, so you used Intel's IPEX-LLM project — a special fork of Ollama and llama.cpp with Intel GPU optimizations. It worked well. The best published B580 numbers (below) were measured on it.

What changed:

• IPEX-LLM is frozen. The GitHub repository was archived on January 28, 2026, with a notice that Intel will not provide or guarantee further development or support. The final portable zips (v2.2.0 stable, v2.3.0-nightly) still download and still run, but they are based on Ollama v0.6.2 — an early-2025 core. Model families released since then will not run on it, and nothing will be fixed again.
• Official Ollama picked up Vulkan. Ollama's GPU documentation now lists Vulkan as additional GPU support on Windows and Linux, "enabled by default when the backend is installed." Intel Arc rides that path with no special fork. You disable it with OLLAMA_VULKAN=0; you select GPUs with GGML_VK_VISIBLE_DEVICES.
• llama.cpp's Vulkan backend grew up. The community scoreboard in the long-running "Performance of llama.cpp with Vulkan" discussion currently records the B580 at 70 tok/s generation and 621 t/s prompt processing on Llama 2 7B Q4_0 — generation speed in the same band as Intel's own IPEX-LLM stack posted on 8B models, from a generic backend that needs no Intel fork at all. The early-2025 refrain that "Vulkan is slow on Arc" does not survive contact with that table.
• Mainline SYCL is not the answer on Battlemage. More on that below — recent llama.cpp issue reports show it broken or far slower than Vulkan on B-series cards.

The takeaway: the B580 went from "works, but only through Intel's special fork" to "works with the standard tools everyone else uses." That is a better place to be, even if it arrived by way of Intel abandoning its own project.

---

Real Published Speeds {#benchmarks}

Answer first: expect roughly 60-80 tok/s on 7-8B Q4 models, ~70 tok/s on 7B via modern Vulkan builds, and 32-38 tok/s on 14B Q4 under Ubuntu. On Windows, 14B can collapse to ~15 tok/s from VRAM pressure.

We have not benchmarked a B580 in-house yet, so every number here is attributed to its source. The two best public data sets:

Three things worth internalizing from that table.

First, the OS matters more than the backend for 14B models. AbelChen's Windows testing found that background apps eat enough of the 12GB that a ~9GB 14B model spills into shared system memory and speed collapses to ~15 tok/s — recoverable to ~40 tok/s only by rebooting and running nothing else. The same card on Ubuntu held 32-38 tok/s with background load. His conclusion, which we endorse: if 14B models are the point, run Linux.

Second, Vulkan generation speed is no longer the compromise — prompt processing is. At 70 tok/s generation, plain Vulkan sits in the same band as IPEX-LLM's 8B figures, which retires the "Arc needs Intel's special fork to be usable" advice from early 2025. But note the A770 contrast row: the older card posts 1,074 t/s prompt processing against the B580's 621 t/s on the same scoreboard. If your workload front-loads huge prompts — big RAG contexts, whole-file coding assistance — the wait before the first token is where the B580 will feel slower than its generation speed suggests.

Third, 12GB draws a hard line at 14B. 8B Q4 models use about 5GB and leave lots of context headroom; 14B Q4 uses about 9GB and fits with care; anything meaningfully bigger spills. Our 12GB VRAM model picks and best Ollama models for 12GB are curated for exactly this card.

---

Setup Path 1: Official Ollama with Vulkan (Recommended) {#ollama-vulkan}

Answer first: install the current Arc driver, install standard Ollama, and run a model — the Vulkan backend is enabled by default. No fork, no container, no oneAPI toolkit.

Windows:

1. Install the current Intel Arc graphics driver from intel.com.
2. Install Ollama from ollama.com (standard installer).
3. Pull and run a model sized for 12GB:

ollama run llama3.1:8b
ollama run qwen2.5-coder:14b

Linux (Ubuntu): the Vulkan path rides the open-source Mesa driver that ships with your distro, and Arc support there improves release by release — so use a current one (Ubuntu 24.04 LTS or newer) rather than resurrecting an old install.

curl -fsSL https://ollama.com/install.sh | sh
ollama run llama3.1:8b

Useful switches, per Ollama's GPU docs:

OLLAMA_VULKAN=0           # disable the Vulkan backend entirely
GGML_VK_VISIBLE_DEVICES=0 # pin which Vulkan device to use

Two practical notes. Enable Resizable BAR in UEFI before anything else — Arc cards are notorious for crawling without it. And verify the model actually landed on the GPU: run ollama ps and check the load shows GPU, not CPU; if it says CPU, your driver or Vulkan runtime is the problem, not Ollama.

---

Setup Path 2: The IPEX-LLM Portable Zip (Frozen, but Fast) {#ipex-portable}

Answer first: the archived IPEX-LLM portable zip still works and still posts the best documented 14B numbers on this card — but it is frozen on an Ollama v0.6.2 core, so models released after early 2025 will not run. Use it only if a model it supports is exactly what you need.

This is the stack AbelChen's 60-80 tok/s numbers were measured on, and since it was a purpose-built Intel GPU runtime rather than a generic backend, it squeezed real performance out of the card. The final builds live on the IPEX-LLM GitHub releases page (tag v2.3.0-nightly).

Windows: download the portable zip, extract, then in Command Prompt:

cd /d PATH\TO\EXTRACTED\FOLDER
start-ollama.bat
ollama run deepseek-r1:7b

Linux: download the tgz, then:

tar -xvf ollama-ipex-llm-*.tgz
cd ollama-ipex-llm-*
./start-ollama.sh
./ollama run deepseek-r1:7b

Environment variables the quickstart documents:

OLLAMA_NUM_CTX=16384                  # bigger context window
ONEAPI_DEVICE_SELECTOR=level_zero:0   # pin a specific Intel GPU
OLLAMA_NUM_PARALLEL=1                 # save VRAM

Be clear-eyed about what "archived" means: no fixes, no new model support, ever. Every month this path gets less useful as the interesting models move past its frozen core. We keep it documented because the performance is real and verified — but if you are setting up fresh in late 2026, start with Path 1 and only reach for this if a model it supports runs poorly on Vulkan for you.

---

llama.cpp Directly: Vulkan Yes, SYCL No {#llama-cpp}

Answer first: use the prebuilt Vulkan binaries from llama.cpp releases. Do not follow older Arc guides down the SYCL build path — recent issue reports show SYCL garbled or dramatically slower than Vulkan on Battlemage cards.

If you want llama.cpp itself (for llama-server, grammars, or first-day GGUF support), the setup is refreshingly boring: download the Vulkan build from the llama.cpp releases page — no compile needed — and run:

llama-server -m your-model.gguf -ngl 99

The reason we are explicit about avoiding SYCL: our Arc A770 guide documented the SYCL build path because it was the right answer for Alchemist cards. On Battlemage the record got ugly. The SYCL docs technically still list the Arc B580 as supported — it builds and runs — but llama.cpp's issue tracker collected reports through spring 2026 of SYCL producing garbled output on Intel GPUs after the first prompt (issues #21589, #21734) and of "brutally bad SYCL performance on Battlemage" (issue #22413 — the reporter's title, not ours), while the same cards behaved on Vulkan. Some of those issues have since been closed, but the community benchmarks, the working setups, and the development attention are all on the Vulkan side. Follow the attention.

---

Stable Diffusion and Flux on the B580 {#stable-diffusion}

Answer first: yes — ComfyUI runs on the B580 through native PyTorch XPU wheels, and 12GB comfortably covers SD 1.5, SDXL, and quantized Flux Schnell. For a zero-setup Windows option, Intel's AI Playground app bundles all three.

The clean path is ComfyUI with PyTorch's native Intel GPU support (no IPEX extension needed anymore — torch.xpu is built into current PyTorch). Per the ComfyUI README:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/xpu
git clone https://github.com/comfyanonymous/ComfyUI
cd ComfyUI && pip install -r requirements.txt
python main.py

We have not measured B580 image-generation times ourselves, so we won't quote seconds-per-image — but on capacity, 12GB puts you in the same class as an RTX 3060 12GB: SD 1.5 and SDXL with headroom, Flux Schnell at reduced precision. Our Stable Diffusion picks for 12GB match models to this card, and the ComfyUI complete guide covers workflows once you're running.

If you would rather skip Python entirely on Windows, Intel AI Playground (Intel's own open-source app, v3.1.2 beta) supports Arc B-series cards with 8GB+ VRAM and ships image generation with Stable Diffusion 1.5, SDXL, and Flux.1-Schnell, plus a local LLM chat with RAG. Linux support exists but Intel labels it experimental.

---

Honest Limits {#limits}

Answer first: the B580 is a 12GB card without CUDA. That means a 14B ceiling for comfortable GPU-only inference, a Linux requirement to actually hit that ceiling, no access to CUDA-only tooling, and a software stack whose vendor-maintained era just ended.

• 12GB is a real ceiling. 14B Q4 fits with ~3GB to spare; 20B+ models spill into system RAM and speeds fall off a cliff. If your ambitions run to 24-32B models, this is the wrong card — see our used GPU buying guide for 16-24GB options at comparable money.
• Windows squanders the VRAM. The published 14B numbers collapse from ~40 to ~15 tok/s under normal Windows background load. This card rewards a dedicated Linux box and punishes a Windows daily driver.
• No CUDA, ever. ExLlama-family loaders, most fine-tuning stacks, and a long tail of research tooling are CUDA-first or CUDA-only. The B580 is an inference card. If fine-tuning is on your roadmap, budget for NVIDIA.
• The vendor stack is gone. IPEX-LLM's archive means Intel no longer maintains a dedicated LLM runtime for its own GPUs. Vulkan support in Ollama and llama.cpp is genuinely good and actively developed — but it is community-maintained, and Arc owners are now downstream of it. Occasional model-specific Vulkan bugs on Arc do get reported; they also get fixed, which was stopped being true of IPEX-LLM.
• PCIe 4.0 x8 and ReBAR. The x8 link is fine for inference, but the card genuinely needs Resizable BAR — budget boards old enough to lack it are a bad pairing.

---

B580 vs A770 vs Used RTX 3060 {#comparison}

Answer first: B580 for speed on models up to 14B, the A770 16GB for capacity, a used RTX 3060 12GB for CUDA compatibility at the same VRAM.

A caveat we insist on: those decode numbers come from different systems, different months, and different test prompts — treat them as bands, not a race result. Directionally, though, the picture is consistent across sources: the newer Xe2 architecture extracts far more LLM performance per GB/s than Alchemist did, the A770's extra 4GB lets it hold 16B-22B models the B580 simply cannot, and the used 3060 remains the compatibility king because CUDA runs everything with zero caveats. Our full A770 guide makes the capacity case if that trade tempts you.

---

Verdict {#verdict}

At or near $249, the Arc B580 is the best new GPU you can buy for local AI, full stop — provided you run Linux and your ambitions top out at 14B models.

Buy it if: you want 7-14B models (chat, coding assistants, RAG) at genuinely usable speeds, you are comfortable with Ubuntu, and you can find the card near MSRP. It is the cheapest new hardware that makes local AI feel fast rather than tolerated.

Skip it if: you need models above 14B (get more VRAM), you need fine-tuning or CUDA-only tools (get NVIDIA), or street price has drifted far above MSRP — at $350+, the calculus shifts toward a used 3060 12GB or saving for a 16GB card.

And whichever way you go, size your models to the card before you buy: our 12GB model picks show exactly what this card runs well.

---

Sources {#sources}

• AbelChen.dev — Intel Arc B580 12GB for Local AI — hands-on Windows-vs-Linux IPEX-LLM benchmarks (Aug 2025)
• llama.cpp discussion #10879 — Performance of llama.cpp with Vulkan — community B580 and A770 llama-bench scoreboard results
• IPEX-LLM repository — archive notice (Jan 28, 2026), portable zip quickstarts (Ollama v0.6.2 base) and releases
• Ollama GPU documentation — Vulkan backend status and environment variables
• llama.cpp SYCL backend docs — B-series support listing
• ComfyUI README — Intel Arc install path via PyTorch XPU wheels
• Intel AI Playground — supported hardware, bundled image models, experimental Linux support
• Intel Arc B580 specifications (Intel ARK) — memory, bandwidth, PCIe 4.0 x8, $249 pricing
• llama.cpp issue tracker — SYCL-on-Battlemage reports, spring 2026 (#21589, #21734, #22413)

---

FAQ {#faq}