Arc B580
An Arc B580 runs 11 of the 34 open models tracked here at Q4_K_M with 8k context, computed on 16 October 2026. The largest is Qwen3 14B, needing 10.8 GB and generating about 37 tokens per second.
- Memory
- 12 GB VRAM
- Bandwidth
- 456 GB/s
- Board power
- 190 W
- Released
- 2024
- Availability
- Sold new
Open the Arc B580 in the LLM checker
What an Arc B580 runs, model by model
Arc B580 (12 GB) addresses 12 GB at 456 GB/s. Every model below is at Q4_K_M with 8k of context, the setting most people actually use. Once the 5% safety margin comes off, the card has 11.4 GB to spend: memory decides what loads, bandwidth how fast it answers. Both are estimates from stated formulas, and the VRAM calculator shows the working.
On this card a model fits at or under 11.4 GB, and above 10.2 GB it is tight: it loads, but there is little left for the conversation, and a desktop on the same 12 GB can push it over. Tokens per second are single-stream estimates from memory bandwidth; ≤ marks a ceiling real runtimes fall well below.
Computed 16 October 2026The model to actually run on it
Gemma 4 12B is the best use of this card: 8.2 GB of the 12 GB available, an estimated 46 tokens per second, comfortable for chat, and 131,072 tokens of context still available.
How this is chosen, since no benchmark is quoted: models are compared by size, a mixture-of-experts model counting at the geometric mean of its total and active parameters (a rule of thumb, not a measurement). The pick is the biggest class that fits with memory to spare and answers at 25 tokens per second or better, and within that class the most recent general-purpose release; coding and reasoning specialists are listed above but not recommended by default.
Qwen3 14B is the largest model the card will hold, at 10.8 GB, but it is the wrong daily driver: filling 90% of memory leaves room for only 11,264 tokens of conversation, and it generates at 37 tokens per second against 46. More parameters are not worth a context window that runs out mid-document.
If quality matters more than parameter count, Qwen3 8B fits at Q8_0 in about 10.7 GB. Q8 is near-lossless; a smaller model at Q8 often beats a larger one squeezed into Q3.
How much context actually fits
Model size is the question people ask; context is the one that bites. The KV cache grows linearly with the conversation, so a model that loads comfortably can still run out of memory halfway through a long document. On 12 GB, the context you get is whatever the weights leave behind, which is why two cards that run the same model can feel completely different in use. These are the largest contexts this card holds, at Q4_K_M weights.
| Model | Max context, FP16 KV | With Q8 KV cache |
|---|---|---|
| Gemma 4 12B | 131,072 tokens | 131,072 tokens |
| Qwen3 14B | 11,264 tokens | 23,552 tokens |
| Phi-4 14B | 9,216 tokens | 16,384 tokens · model limit |
| Ministral 3 14B | 14,336 tokens | 29,696 tokens |
Quantising the KV cache to Q8 roughly doubles what you can hold, at a quality cost most people never notice: Qwen3 14B on this card goes from 11,264 to 23,552 tokens. Capped at 128k tokens, or at the model's own native window where that is smaller, marked "model limit": free memory beyond that point buys nothing. Rows that hold far more context than their size suggests are hybrid, sliding-window or latent-attention models, which cache only a fraction of what a standard transformer does; each model page shows the working.
Where this card stops
The first model out of reach is gpt-oss 20B: about 13.4 GB at Q4_K_M and 8k context, against 12 GB of usable memory. Dropping to Q3_K_M would need about 11 GB, which fits, though Q3 loses enough quality that a smaller model at Q4 is usually the better trade. Offloading the remainder to system RAM works and is ten to fifty times slower; it is a way to see a model run, not a way to use one.
Buy or rent
A reasonable new 12 GB card if your tools are llama.cpp or Ollama and you accept checking each one for Intel support. If CUDA matters, a used RTX 3060 12 GB holds the same models with wider software support, at a lower speed ceiling. For anything above 12 GB, look at the 24 GB tier or rent by the hour.
Retail prices drift, and a figure written today would be wrong within a month, so this page does not carry one. Put the figure you are actually looking at, with your electricity rate and your real monthly hours, into the build vs rent calculator, which already knows this card draws 190 W under load. It returns the month owning becomes cheaper, or tells you it never does. Hours per month decides it far more often than the price does.
The honest take
Intel's 12 GB card: faster memory than an RTX 3060 12 GB, and a younger software stack.
It holds the same models as any 12 GB card and reads them at 456 GB/s, about a quarter faster than an RTX 3060 and a little under an RTX 4070. The hardware case is sound; the software is the question. llama.cpp supports it through SYCL and Vulkan, and Ollama through Vulkan, but Intel has archived its own IPEX-LLM and Extension for PyTorch projects, and CUDA-only tools do not run.
What it is good at
- 456 GB/s on 12 GB: about a quarter more than an RTX 3060 12 GB.
- llama.cpp's SYCL backend lists the B580 as a verified device, and supports Linux and Windows 11.
- A 190 W board with one eight-pin power connector, two slots on Intel's own card.
What it is not
- Twelve gigabytes is the same ceiling as an RTX 3060 or 4070.
- Intel has archived IPEX-LLM and the Intel Extension for PyTorch; Intel now points to upstream PyTorch, and local chat runs through llama.cpp or Vulkan.
- No CUDA: fine-tuning and image tools that ship only CUDA builds do not run.
The thing people get wrong: Intel's spec page still lists "Intel Extension for PyTorch (IPEX) Support: Yes", but Intel has archived that project and IPEX-LLM, and tells users to use PyTorch directly. Guides that start by installing IPEX-LLM now lead to unmaintained software.
Compared with the alternatives
- RTX 3060: 12 GB · 360 GB/s · best fit Gemma 4 12B at ~36 tok/s. The cheapest card that still makes local models worth doing.
- RTX 4070: 12 GB · 504 GB/s · best fit Gemma 4 12B at ~51 tok/s. Fast for its class, and capped by 12 GB.
- Arc Pro B60: 24 GB · 456 GB/s · best fit Gemma 4 12B at ~46 tok/s. Intel's 24 GB workstation card: the B580's bandwidth with twice the memory.
All 39 cards are compared side by side on the GPU index, and the same numbers from the model's point of view are on each model fit table. To check one specific pairing rather than read a table, use the can-I-run-it checker.
Questions
- Can an Arc B580 run a 70B model?
- Not at Q4_K_M in one card. Llama 3.3 70B needs about 45.8 GB at 8k context and this card can address 12 GB. Your options are a smaller model, a harsher quantisation with very little context, splitting across two cards, or renting a larger one for the hours you need it.
- What is the best model to run on an Arc B580?
- Gemma 4 12B. At Q4_K_M and 8k context it needs about 8.2 GB of the 12 GB available and generates an estimated 46 tokens per second, which is comfortable for chat. If quality matters more than size, Qwen3 8B fits at Q8 in about 10.7 GB.
- How many tokens per second does an Arc B580 generate?
- It depends entirely on the size of the model, because generating a token means reading every active weight out of memory. At 456 GB/s and 70% efficiency, this card produces an estimated 69 tokens per second on an 8B model at Q4, and about 37 on the largest model it holds, Qwen3 14B. Reading speed is roughly 5 to 8 words per second, so anything above 25 feels immediate. Treat 69 as a ceiling on 456 GB/s rather than a measurement; the higher a figure, the further real runtimes fall below it.
- Is 12 GB enough for running LLMs locally?
- It runs 11 of the 34 models tracked here at Q4_K_M with 8k of context, up to 14.8B parameters. The honest test is not the model list but the context: Gemma 4 12B on this card holds about 131,072 tokens before memory runs out.
- Arc B580 or RTX 3060 for local models?
- Both address about the same memory, so they run the same models. On speed, this card is faster: 456 GB/s against 360 GB/s, and bandwidth is what sets chat speed.
- Arc B580 or RTX 3060 12 GB for AI?
- Both hold 12 GB, so the same models fit. The B580 reads memory at 456 GB/s against 360 GB/s, about a quarter faster, so its speed ceiling is higher on every model both can hold. The 3060 runs CUDA, which most AI tools support first; the B580 relies on SYCL or Vulkan builds. For llama.cpp or Ollama chat, the B580; for the widest tool support, the 3060.
- Does Ollama support the Intel Arc B580?
- Through Vulkan, yes. Ollama's hardware page says Vulkan adds GPU support on Windows and Linux and is on by default, and it links Intel's Linux driver instructions; it does not list Intel cards by name. The Ollama build from Intel's IPEX-LLM project is archived, so use Ollama's own releases.
- Memory: weights (parameters × bytes per parameter) + KV cache (2 × layers × KV heads × head dim × context × bytes for a standard transformer; sliding-window, hybrid and latent-attention models counted as they actually cache) + 0.5 GB runtime + 4% of weights. Full derivation in the VRAM calculator.
- Speed: 70% × bandwidth ÷ active weight bytes, single stream, no batching, prompt processing excluded. A ceiling, not a measurement: the 2 of 11 figures marked ≤ are where real runtimes fall furthest below the number, because so few weights are read per token. See the tokens-per-second estimator.
- Card specification (manufacturer figures): 12 GB, 456 GB/s, 190 W; see Intel's own page for this card.
- Model architecture values come from each model's published config; the self-hosted LLM guide explains what each one changes.