16GB VRAM

Arc A770 16G — what LLMs can it run?

60 of 112 indexed models fit comfortably in 16GB at 4K context, each at the highest-quality quant that still leaves headroom.

The short answer for Arc A770 16G

16 GB, 560 GB/s GDDR6. 60 of 112 models in this index fit comfortably at 4K context; 69 load at all.

Biggest that fits
ERNIE 4.5 21B-A3B — 14.1 GB at Q4_K_M — 1.9 GB spare at 4K, so longer context comes out of a thin margin.
Room to grow
Stable LM 2 12B Chat — 8.3 GB at Q4_K_M — under 60% of the card, which leaves 7.7 GB for a long context window or a second process.
Speed ceiling
Llama 3.1 8B Instruct at Q4_K_M reads 4.6 GB of weights per token, so 560 GB/s puts a hard ceiling near 121 tok/s. That is arithmetic on two published numbers, not a benchmark — real throughput lands below it. No run on this card has been measured here, so there is nothing to compare the ceiling against.
The wall you will hit
Backend coverage, not capacity. At 560 GB/s the card has the memory system for its size class; what limits it is which runtimes reach it. llama.cpp's SYCL backend lists Arc A-series and B580 cards as verified, and Ollama runs them through Vulkan — so GGUF is the format this site recommends here. AWQ, GPTQ and EXL2 are not.
Best local LLM for 16GB

32B · 2

Arc A770 16G — what LLMs can it run? — 32B
ModelQuantEst. VRAMHeadroomtok/s
ERNIE 4.5 21B-A3B21B-A3BQ4_K_M14.05 GB+1.9 GB≤324Ceiling
GPT-OSS 20B21B MoEMXFP412.76 GB+3.2 GB≤283Ceiling

14B · 19

Arc A770 16G — what LLMs can it run? — 14B
ModelQuantEst. VRAMHeadroomtok/s
InternLM2 20B Chat20BQ4_K_M13.43 GB+2.6 GB17–22Derived
DeepSeek-Coder-V2-Lite Instruct16B-A2.4BQ4_K_M10.08 GB+5.9 GB≤405Ceiling
DeepSeek-V2-Lite Chat16B-A2.4BQ4_K_M10.08 GB+5.9 GB≤405Ceiling
StarCoder2 15B15BQ4_K_M10.19 GB+5.8 GB22–29Derived
Qwen3 14B Instruct14BQ5_K_M11.65 GB+4.4 GB19–25Derived
Qwen2.5 14B Instruct14BQ5_K_M11.73 GB+4.3 GB19–25Derived
DeepSeek-R1-Distill-Qwen-14B14BQ4_K_M10.14 GB+5.9 GB22–30Derived
Phi-4 14B14BQ5_K_M11.77 GB+4.2 GB19–25Derived
Phi-3 Medium 14B Instruct14BQ6_K12.86 GB+3.1 GB17–23Derived
Mistral Nemo 12B Instruct12BQ6_K11.14 GB+4.9 GB20–26Derived
Gemma 3 12B IT12BQ5_K_M9.84 GB+6.2 GB23–30Derived
Mellum2.1 12B-A2.5B Thinking12B-A2.5BQ8_013.6 GB+2.4 GB≤222Ceiling
Stable LM 2 12B Chat12BQ4_K_M8.35 GB+7.7 GB27–36Derived
Gemma 4 12B IT12BQ8_013.82 GB+2.2 GB16–21Derived
Llama 3.2 11B Vision Instruct11BQ8_012.9 GB+3.1 GB17–23Derived
Solar 10.7B Instruct11BQ4_K_M7.6 GB+8.4 GB31–41Derived
Falcon 3 10B Instruct10BQ4_K_M7.28 GB+8.7 GB32–42Derived
Gemma 2 9B Instruct9BQ8_011.7 GB+4.3 GB20–27Derived
Qwen3.5 9B9BQ8_010.37 GB+5.6 GB20–27Derived

7B · 25

Arc A770 16G — what LLMs can it run? — 7B
ModelQuantEst. VRAMHeadroomtok/s
GLM-4-9B-Chat9BQ8_010.16 GB+5.8 GB21–28Derived
Qwen3-VL 8B Instruct8BQ8_010.39 GB+5.6 GB21–28Derived
Granite 4.2 8B8BQ8_010.45 GB+5.6 GB21–28Derived
Ministral 3 8B Instruct8BQ8_010.02 GB+6 GB22–29Derived
Qwen2-VL 7B Instruct7BQ4_K_M5.49 GB+10.5 GB40–52Derived
Granite 3.1 8B Instruct8BQ4_K_M5.88 GB+10.1 GB40–53Derived
Qwen3 8B Instruct8BQ6_K7.64 GB+8.4 GB30–39Derived
Llama 3.1 8B Instruct8BQ8_09.47 GB+6.5 GB23–31Derived
Nous Hermes 3 Llama 3.1 8B8BQ4_K_M5.64 GB+10.4 GB41–54Derived
Aya 23 8B8BQ4_K_M5.64 GB+10.4 GB41–54Derived
OpenChat 3.6 8B8BQ4_K_M5.64 GB+10.4 GB41–54Derived
DeepSeek-R1-Distill-Llama-8B8BQ5_K_M6.51 GB+9.5 GB35–46Derived
InternLM2 7B Chat7BQ4_K_M5.45 GB+10.6 GB42–56Derived
Qwen2.5 7B Instruct7BQ6_K6.77 GB+9.2 GB32–42Derived
Qwen2.5-Coder 7B Instruct7BQ4_K_M5.07 GB+10.9 GB43–57Derived
DeepSeek-R1-Distill-Qwen-7B7BQ4_K_M5.07 GB+10.9 GB43–57Derived
Gemma 4 E4B ITE4BQ8_08.46 GB+7.5 GB25–33Derived
OLMo 2 7B Instruct7BQ8_010.3 GB+5.7 GB26–34Derived
OLMo 3 7B Instruct7BQ8_010.3 GB+5.7 GB26–34Derived
WizardLM-2 7B7BQ4_K_M5.14 GB+10.9 GB45–60Derived
Mistral 7B Instruct v0.37BQ6_K6.75 GB+9.3 GB34–44Derived
Zephyr 7B Beta7BQ6_K6.75 GB+9.3 GB34–44Derived
Gemma 4 E2B ITE2BQ8_05.2 GB+10.8 GB≤119Ceiling
Qwen3.5 4B4BQ8_04.97 GB+11 GB≤128Ceiling
Gemma 3 4B IT4BQ8_05.05 GB+11 GB≤129Ceiling

≤3B · 14

Arc A770 16G — what LLMs can it run? — ≤3B
ModelQuantEst. VRAMHeadroomtok/s
Qwen3 4B Instruct4BQ6_K4.05 GB+12 GB≤179Ceiling
Phi-4 Mini Instruct3.8BQ8_04.81 GB+11.2 GB≤144Ceiling
Phi-3.5 Mini Instruct3.8BQ8_05.89 GB+10.1 GB≤145Ceiling
Granite 4.2 3B3BQ8_04.41 GB+11.6 GB≤152Ceiling
Llama 3.2 3B Instruct3BQ8_04.05 GB+12 GB≤173Ceiling
Qwen2.5 3B Instruct3BQ8_03.59 GB+12.4 GB≤179Ceiling
SmolLM3 3B3BQ8_03.73 GB+12.3 GB≤180Ceiling
Gemma 2 2B Instruct2BQ8_03.34 GB+12.7 GB≤213Ceiling
Qwen3.5 2B2BQ8_02.23 GB+13.8 GB≤283Ceiling
Qwen3 1.7B Instruct1.7BQ8_02.39 GB+13.6 GB≤322Ceiling
Qwen2.5 1.5B Instruct1.5BQ8_01.83 GB+14.2 GB≤361Ceiling
Llama 3.2 1B Instruct1BQ8_01.51 GB+14.5 GB≤448Ceiling
Qwen3.5 0.8B0.8BQ8_00.94 GB+15.1 GB≤691Ceiling
Qwen2.5 0.5B Instruct0.5BQ8_00.6 GB+15.4 GB≤1143Ceiling

How this list is built

Each row is the lowest-perplexity-loss quant of that model whose estimated total — weights plus KV cache at 4K context plus activation buffer — uses at most 88% of the card. That is the calculator's "green" threshold, so every row here has real headroom rather than only just fitting. Raise the context length and the list shortens; the calculator lets you check any combination directly. The speed column is this card: a measured run, a range scaled from this card's own Llama 2 7B benchmark, or a bandwidth ceiling. A figure that only exists for the RTX 4090 is labelled as that card.

A further 9 models load but with no headroom to spare (up to 105% of VRAM) — the Quant Hub’s GPU chips count those too, which is why its number is higher.

Open this card in the VRAM calculator Feel this speed Browse all 112 models

Measured on this card

No benchmark runs in this index were recorded on a Arc A770 16G. Every figure on this page is calculated from the model architecture and the quant level — treat them as estimates, not measurements.

Cards with the same budget

The same 16GB in system memory is a different machine — everything on this page fits there, at a speed set by your DIMMs rather than a graphics bus: 16 GB RAM (CPU)

    Setup guides for this card

    Step-by-step guides written for the Arc A770 16G's runtime or memory size, with the commands and the checks that prove the model is running on the card.

    Stepping up

    A Arc Pro B70 32G (32GB) fits 28 more of the indexed models than this card. Arc Pro B70 32G →

    Before buying, hear the speed difference on a model both cards hold: Arc A770 16G vs Arc Pro B70 32G (Qwen3 14B Instruct) →

    Common questions

    What is the best local LLM for an Arc A770 16G?

    For everyday use, Stable LM 2 12B Chat at Q4_K_M — about 8.3 GB of the card's 16 GB at 4K context, leaving 7.7 GB for a longer window. If you want the largest thing that will load, that is ERNIE 4.5 21B-A3B at Q4_K_M (14.1 GB, 1.9 GB spare). "Best" here means best fit for the memory budget — this index does not run task benchmarks, so it cannot tell you which model is smarter.

    Can an Arc A770 16G run Codestral 22B?

    Only just. Its smallest build here, Q4_K_M, needs about 15.0 GB against 16.0 GB usable — that loads with nothing else running, with no margin for a longer context window. It is not on the list above, which requires a model to stay inside 88% of usable capacity.

    How many tokens per second does an Arc A770 16G do on an 8B model at Q4?

    This index has no measured run on this card, so it does not publish a figure. What can be stated from specifications: generating a token requires reading every weight once, Llama 3.1 8B Instruct at Q4_K_M is 4.6 GB of weights, and this card moves 560 GB/s — a ceiling near 121 tok/s. Batch size, context length, the runtime and how much of the model sits in cache all take you below it.

    Arc A770 16G or RTX 5080 for local LLMs?

    They do not hold the same set: Arc A770 16G fits 60 and RTX 5080 64 of 112 at 4K — the cards load different formats. Speed is the other difference — 560 GB/s against 960 GB/s, a 1.7× gap in how fast the weights can be read, which is what token generation is bound by. The RTX 5080 generates faster on any model both can hold.

    60 of 112 indexed models fit comfortably in 16GB at 4K context, each at the highest-quality quant that still leaves headroom.

    This page's figures change when the model or the runtime does.Last updated 2026-10-20 RSS → /feed.xml