Home/Hardware/Intel Arc A770
Intel Arc A770 · 16GB GDDR6 on 560 GB/s — Vulkan or SYCL, no CUDA

Best AI Models for the Intel Arc A770

The full 16GB model list on a card that sells used for about $350, and the exact state of Ollama, llama.cpp and Intel’s own tools on it. Footprints from Ollama’s tag sizes, speeds as bandwidth ceilings.

📅 Published: November 20, 2026🔄 Last Updated: November 2026✓ Manually Reviewed
Short answer

The best AI model for the Intel Arc A770 is Qwen 3 14B — 9.3GB at Q4_K_M as Ollama lists it, leaving ~6–7GB of the 16GB for context, with a bandwidth ceiling of ~60 tok/s (Intel’s 560 GB/s ÷ 9.3GB — a ceiling, not a measurement). Fit is decided by memory, not by vendor, so the A770 holds the same list as an RTX 5070 Ti: Gemma 4 12B at Q8 (13GB) for near-lossless quality, Qwen 2.5 Coder 14B (9.0GB) for code and gpt-oss:20b (14GB, MoE) as the fast assistant. The catch is software. Ollama has no Intel backend of its own: its GPU docs name no Arc card, and Intel GPUs fall under its Vulkan backend. The Intel-native route is llama.cpp’s SYCL backend, which lists the A770 as a verified device. Intel’s IPEX-LLM, the route many A770 guides still describe, has been archived by Intel. We do not own an A770; nothing here is a measured rate.

Models that run well on the Intel Arc A770

ModelSizeBuildMemory usedCeiling tok/s*
WINNERQwen 3 14B
The 16GB winner on any vendor’s card: the whole 14B class with ~6–7GB of KV budget left over. Its q8_0 tag is 16GB, the entire card, so Q4 is the build to run.
ollama pull qwen3:14b
14B (dense)Q4_K_M~9.3GBceiling ~60 tok/s (560 ÷ 9.3GB)
Gemma 4 12B
The quality lane: Google’s dense 12B at Q8 with text and image input and ~3GB spare. The q4_K_M build is 8.0GB if you would rather spend the memory on context.
ollama pull gemma4:12b-it-q8_0
12B (dense)Q8_0~13GBceiling ~43 tok/s (560 ÷ 13GB)
Qwen 2.5 Coder 14B
The coding pick with room to work, ~7GB of KV budget for whole files. A GGUF file is vendor-neutral, so the same download runs on Vulkan, SYCL or CUDA.
ollama pull qwen2.5-coder:14b
14B (dense)Q4_K_M~9.0GBceiling ~62 tok/s (560 ÷ 9.0GB)
gpt-oss:20b
OpenAI’s open-weight model, sized for 16GB cards. It fills most of the A770, so run it alone with a modest context.
ollama pull gpt-oss:20b
20.9B (MoE, ~3.6B active)MXFP4~14GBMoE — ~3.6B active per token; bandwidth is not the limit; not measured
Devstral Small 2 24B
The strongest agentic coder that fits, with about a gigabyte left for the KV cache. It also has the lowest ceiling on this list, which suits batch refactors better than interactive autocomplete.
ollama pull devstral-small-2:24b
24B (dense)Q4_K_M~15GBceiling ~37 tok/s (560 ÷ 15GB)
SDXL 1.0
Image generation goes through ComfyUI on PyTorch’s own XPU build, which PyTorch validates on Arc A-Series cards. Guides that install Intel’s separate PyTorch extension (IPEX) are out of date: Intel has archived it.
ComfyUI on PyTorch XPU (download.pytorch.org/whl/xpu)
3.5BFP16~7GBdiffusion — compute-bound, not timed

* Ceiling tok/s = the device’s vendor-listed memory bandwidth ÷ the model’s file size as Ollama lists it — the rate a dense model cannot exceed, because every generated token reads the whole weight set once. We have not measured this hardware. Real-world generation usually lands at 50–80% of the ceiling and falls as the context grows; mixture-of-experts rows read only their active parameters, so no ceiling is quoted for them, and image models are compute-bound, so no render time is.

What won't fit on the Intel Arc A770

  • ✗Qwen3.6 27B (needs 17GB at Q4_K_M) — One gigabyte past the card before any context. The 27B class is what 24GB cards are for; on 16GB, the 14B class is the ceiling.
  • ✗Qwen 3 14B at Q8_0 (needs 16GB) — The tag is the size of the whole card. Near-lossless on 16GB means a 12B (Gemma 4 12B, 13GB).
  • ✗Qwen3-Coder 30B (needs 19GB at Q4_K_M) — An MoE that would decode fast here if it fit; it misses by 3GB. Devstral Small 2 is the coding ceiling on 16GB.
  • ✗Llama 3.3 70B (needs 43GB at Q4_K_M, 26GB even at Q2_K) — No 16GB card holds a 70B at any usable quant.

Intel Arc A770-specific notes

  • →Where this page’s numbers come from: we do not own an A770. Footprints are the file sizes Ollama prints on each tag page (read 20 Nov 2026), and every tok/s is a ceiling — the 560 GB/s on Intel’s spec page divided by the gigabytes a dense model reads per token. The 50–80% rule of thumb in the table footnote comes from CUDA and Metal. llama.cpp’s SYCL documentation publishes its own A770 results, and even its improved 2025 figure is under half the ceiling for the model it used. On that evidence, expect a smaller share of the ceiling on this card than on an NVIDIA one, and measure your own setup before trusting anyone’s number.
  • →Ollama support, stated exactly: Ollama’s hardware page (docs.ollama.com/gpu) has named support tables for NVIDIA (CUDA) and AMD (ROCm) and none for Intel. Intel GPUs come under its Vulkan section, “Additional GPU support on Windows and Linux is provided via Vulkan”, which is “enabled by default when the backend is installed”. On Windows the docs say most vendors’ drivers bundle Vulkan; on Linux they point to Intel’s client GPU driver instructions. So Ollama can use an A770 through Vulkan, but no Arc card appears on a tested list. Load a model and run `ollama ps`: 100% GPU means the card was picked up.
  • →A Linux message to ignore: on a machine whose only GPU is an Arc, Ollama’s install script finishes with “No NVIDIA/AMD GPU detected. Ollama will run in CPU-only mode.” The script only looks for NVIDIA and AMD hardware; the Vulkan libraries still ship in Ollama’s standard Linux package, so trust `ollama ps`, not that line. For accurate scheduling under Vulkan, Ollama’s docs also ask for root or `sudo setcap cap_perfmon+ep /usr/local/bin/ollama`; without it Ollama plans around approximate model sizes instead of the card’s free VRAM.
  • →If Vulkan will not cooperate, the Intel-native path is llama.cpp built with its SYCL backend: the SYCL documentation lists the Arc A770 as a verified device on Linux and Windows and builds against Intel’s oneAPI toolkit (run `source /opt/intel/oneapi/setvars.sh` in every shell first). On Windows, Intel’s open-source AI Playground app is the no-terminal option: it runs chat models through llama.cpp’s Vulkan backend or OpenVINO and supports Arc A-Series and B-Series cards with 8GB or more. Our A770 setup guide covers drivers and the SYCL build.
  • →The advice to skip: many A770 guides still send you to Intel’s IPEX-LLM, either its Ollama container or its portable zip. Intel has archived that project; its README now says Intel will not provide maintenance, bug fixes or updates, and that the project “has been identified as having known security issues”. Intel’s PyTorch extension (IPEX) is archived as well. Use upstream Ollama or llama.cpp for language models and PyTorch’s own XPU build for ComfyUI.
  • →The 8GB trap: Intel also sold an Arc A770 with 8GB (16 Gbps memory, 512 GB/s on Intel’s spec page). It loses every row above 8GB here — gpt-oss:20b, Devstral, Gemma 4 at Q8 — and the whole 14B class, which needs 9GB before any context. On 20 Nov 2026 resaleprices.com put the used 8GB card’s average at $355 against $349 for the 16GB, so the smaller card is not even the cheaper one; look for “16GB” in the listing. The laptop chip, the Arc A770M, keeps 16GB (512 GB/s, 120–150 W per Intel), so this list applies to it at a lower ceiling.
  • →Check the platform before judging speed: Intel’s support article says Resizable BAR “must be enabled for optimal performance” on Arc A-Series cards, though they still work without it, so turn on Above 4G Decoding and Re-Size BAR in the UEFI. Intel’s spec page lists 225 W total board power and a PCIe 4.0 x16 interface; power-connector layouts vary by board partner, so check the specific card.
  • →Price on 20 Nov 2026: we found no new-card tracker for the A770, so this is the used market. resaleprices.com put the 16GB card at a $349 average, with a fair asking range of $300–360 from 19 eBay listings. The same tracker’s averages were $683 for the RTX 5060 Ti (a pool of 8GB and 16GB cards) and $889 for the RX 9070 XT. On bandwidth the cheapest of the three sits in the middle: 560 GB/s, against 448 for the 5060 Ti and 640 for the 9070 XT. What the low price does not buy is CUDA: ComfyUI custom nodes with compiled CUDA kernels, most training scripts and TensorRT are off the table.

Quick start

Linux: install Intel’s client GPU driver (dgpu-docs.intel.com), then check the A770 shows up as a Vulkan device
vulkaninfo --summary
Install Ollama (Arc runs on its Vulkan backend; on an Intel-only Linux box, ignore the installer’s CPU-only warning)
curl -fsSL https://ollama.com/install.sh | sh
Linux: let Ollama read free VRAM under Vulkan, then restart the ollama service (from Ollama’s GPU docs)
sudo setcap cap_perfmon+ep /usr/local/bin/ollama
Run the winner
ollama run qwen3:14b
Confirm it is fully on the GPU (must read 100% GPU)
ollama ps
Alternative: llama.cpp with the SYCL backend (run source /opt/intel/oneapi/setvars.sh first)
cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx && cmake --build build --config Release -j
Once your hardware is sorted

Go from "it runs" to actually building

All 561 chapters — running local models, RAG, agents, fine-tuning — plus the Python Lab and every course added later.

$149 once unlocks everything, forever — about $0.27/chapter for life. Prefer to spread it out? Pro is $79/year (saves 27%) or $8.99/month.
Secure checkout by Lemon Squeezy — your card never touches this siteInstant access the moment you payFirst chapter of every course is free — try before you buy

Frequently asked questions

Can Ollama use an Intel Arc A770?

Yes, through its Vulkan backend, with a caveat worth knowing. Ollama’s GPU documentation keeps named support lists for NVIDIA and AMD only; Intel cards fall under “Additional GPU support on Windows and Linux is provided via Vulkan”, which is on by default. No Arc model is named, so think of it as supported by the backend rather than tested by name. Load a model and run `ollama ps`: 100% GPU means it worked. If it shows CPU, the Intel-native alternative is llama.cpp with the SYCL backend, whose documentation lists the A770 as verified. Intel’s IPEX-LLM build of Ollama is archived and should not be used for a new setup.

What is the best LLM for the Arc A770 16GB?

Qwen 3 14B at Q4_K_M — 9.3GB as Ollama lists it, ~6–7GB of the card left for context, and a bandwidth ceiling of ~60 tok/s (560 ÷ 9.3GB). For code, Qwen 2.5 Coder 14B (9.0GB) as the daily driver or Devstral Small 2 24B (15GB) when quality matters more than context; for near-lossless general use, Gemma 4 12B at Q8 (13GB); for a fast assistant, gpt-oss:20b (14GB). Every figure is a tag size or a ceiling, not a rate we measured.

How many tokens per second does an Arc A770 get?

We have not measured one, so we will not print a number as if we had. The ceiling is arithmetic: Intel lists 560 GB/s of memory bandwidth, and a dense model reads its whole weight set once per token, so a 9.3GB model cannot exceed ~60 tok/s and a 13GB one ~43. How close you get depends on the backend — Vulkan and SYCL are separate code paths that still change release by release — and llama.cpp’s own published A770 results sit under half the ceiling. A figure that does not name the backend, build, model file and context length tells you little.

Is the Arc A770 still worth buying for local AI?

For the right buyer. At about $349 used (resaleprices.com, 20 Nov 2026) it is the cheapest route into 16GB of any card on this hub, and 16GB is what holds the 14B class with real context, gpt-oss:20b and Devstral Small 2. The costs are all software: no CUDA, Ollama support through Vulkan rather than a named list, and Intel’s own LLM stack archived. If your tools are Ollama, llama.cpp and ComfyUI, it works. If you rely on CUDA-only projects, it is the wrong card at any price.

Arc A770 8GB vs 16GB for AI?

Only the 16GB card makes sense for this. Intel’s spec pages list the 8GB A770 at 512 GB/s against the 16GB card’s 560, but the memory is what decides the list: the 8GB card cannot hold a 14B model (9.0–9.3GB at Q4), gpt-oss:20b, Devstral Small 2 or any Q8 row, which leaves it in the 7–9B class. On 20 Nov 2026 the used 8GB card averaged $355 on resaleprices.com against $349 for the 16GB, so there is no saving to trade for it.

Related guides

Other GPUs

Ready to Go Beyond Tutorials?

25 structured courses with hands-on chapters - build RAG chatbots, AI agents, and ML pipelines on your own hardware.

Was this helpful?

Free Tools & Calculators