Skip to content
32GB VRAM

Arc Pro B70 32G — what LLMs can it run?

123 of 151 indexed models fit comfortably in 32GB at 4K context, each at the highest-quality quant that still leaves headroom.

Context length
Concurrency (batch)

The short answer for Arc Pro B70 32G

32 GB, 608 GB/s GDDR6. 123 of 151 models in this index fit comfortably at 4K context; 126 load at all.

Biggest that fits
Kolibri-1 78B-A3.5B — 25.6 GB at IQ2_XS — 6.4 GB spare at 4K, so longer context comes out of a thin margin.Download: bartowski/Kolibri-1-GGUF · Kolibri-1-IQ2_XS.gguf · 24.34 GB · listing read 2026-11-12
Room to grow
Gemma 3 27B IT — 18.3 GB at Q4_K_M — under 60% of the card, which leaves 13.7 GB for a long context window or a second process.Download: bartowski/google_gemma-3-27b-it-GGUF · google_gemma-3-27b-it-Q4_K_M.gguf · 16.55 GB · + vision projector 0.86 GB, separate · listing read 2026-11-03
Speed ceiling
Llama 3.1 8B Instruct at Q4_K_M reads 4.6 GB of weights per token, so 608 GB/s puts a hard ceiling near 132 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 608 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 32GB

50B+ · 3

Arc Pro B70 32G — what LLMs can it run? — 50B+
ModelQuantEst. VRAMHeadroomtok/s
Kolibri-1 78B-A3.5B78B-A3.5BPair fit detailsIQ2_XS25.59 GB+6.4 GB≤588Ceiling
Nemotron 3.5 Bolt 74B-A9B74B-A9BPair fit detailsIQ2_XXS23.05 GB+9 GB≤239Ceiling
Ripple-65B-Preview65BPair fit detailsIQ2_S27.18 GB+4.8 GB≤25Ceiling

20–50B · 42

Arc Pro B70 32G — what LLMs can it run? — 20–50B
ModelQuantEst. VRAMHeadroomtok/s
Recursive Flint 40B-A2.5B40B-A2.5BPair fit detailsQ4_K_M26.68 GB+5.3 GB≤402Ceiling
K2 Horizon MoVA 36B-A4B36B-A4BPair fit detailsQ4_K_M24.21 GB+7.8 GB≤268Ceiling
Seed-OSS 36B Instruct36BPair fit detailsQ5_K_M27.83 GB+4.2 GB≤25Ceiling
Qwen3.6 35B-A3B35B-A3BPair fit detailsQ4_K_M23.38 GB+8.6 GB≤340Ceiling
Apex 35B V1.035B-A3BPair fit detailsQ4_K_M23.91 GB+8.1 GB≤332Ceiling
Command R 35B35BQ4_K_M22.86 GB+9.1 GB≤30Ceiling
K2 Horizon 32B32BPair fit detailsQ5_K_M26.91 GB+5.1 GB≤26Ceiling
MiniCPM-V 4.7 35B-A3B35B-A3BPair fit detailsQ5_K_M25.91 GB+6.1 GB≤299Ceiling
Clover-35B-A3B-Preview35B-A3BPair fit detailsQ4_K_M23.4 GB+8.6 GB≤331Ceiling
Yi 1.5 34B Chat34BQ4_K_M22.82 GB+9.2 GB≤31Ceiling
Nanbeige4.3 33B33BPair fit detailsQ4_K_M22.61 GB+9.4 GB≤31Ceiling
Qwen3 32B Instruct32BQ4_K_M21.85 GB+10.1 GB21–28Derived
Qwen2.5 32B Instruct32BQ4_K_M21.69 GB+10.3 GB22–28Derived
Qwen2.5-Coder 32B Instruct32BQ4_K_M21.69 GB+10.3 GB22–28Derived
DeepSeek-R1-Distill-Qwen-32B32BQ4_K_M21.69 GB+10.3 GB22–28Derived
Nemotron 3.5 Lightning 30B-A3B30B-A3BPair fit detailsQ4_019.79 GB+12.2 GB≤356Ceiling
Xing4.0 29B-A4B29B-A4BPair fit detailsQ5_K_M25.28 GB+6.7 GB≤208Ceiling
Gemma 4 31B IT31BPair fit detailsQ5_K_M25.25 GB+6.8 GB19–25Derived
Qwen3-Coder 30B-A3B Instruct30B-A3BPair fit detailsQ5_K_M23.12 GB+8.9 GB≤300Ceiling
Qwen3-VL 30B-A3B Instruct30B-A3BPair fit detailsQ4_K_M19.89 GB+12.1 GB≤350Ceiling
Qwen3 30B-A3B Instruct30B-A3BQ5_K_M23.04 GB+9 GB≤301Ceiling
North Mini Code 1.0 30B-A3B30B-A3BPair fit detailsQ4_K_M20.01 GB+12 GB≤347Ceiling
Hy-Memory-30B-A3B30B-A3BPair fit detailsQ6_K26.21 GB+5.8 GB≤260Ceiling
GLM-4.7-Flash30B-A3BPair fit detailsQ4_K_M19.51 GB+12.5 GB≤346Ceiling
Granite 4.2 30B30BPair fit detailsQ6_K26.19 GB+5.8 GB18–23Derived
Muse Glimmer 30B30BPair fit detailsQ6_K24.61 GB+7.4 GB18–24Derived
Qwen3.8 27B27BPair fit detailsQ6_K25.2 GB+6.8 GB18–24Derived
Arete 27B27BPair fit detailsQ5_K_M21.89 GB+10.1 GB21–27Derived
Qwen3.6 27B27BPair fit detailsQ5_K_M22.19 GB+9.8 GB20–27Derived
DaisyRose 27B27BPair fit detailsQ5_K_M22 GB+10 GB20–27Derived
Gemma 2 27B Instruct27BQ5_K_M21.76 GB+10.2 GB22–29Derived
Gemma 3 27B IT27BPair fit detailsQ4_K_M18.3 GB+13.7 GB26–34Derived
Ternary Bonsai 2 27B27BPair fit detailsQ2_08.24 GB+23.8 GB56–74Derived
Clairic 27B27BPair fit detailsQ5_K_M20.51 GB+11.5 GB22–29Derived
Gemma 4 26B-A4B IT26B-A4BPair fit detailsQ4_K_M18.22 GB+13.8 GB≤237Ceiling
Mistral Small 24B Instruct24BQ4_K_M15.89 GB+16.1 GB29–39Derived
LFM2.5 24B-A2B24B-A2BPair fit detailsQ8_026.73 GB+5.3 GB≤246Ceiling
Devstral Small 1.1 24B24BPair fit detailsQ6_K20.89 GB+11.1 GB22–29Derived
Magistral Small 1.2 24B24BPair fit detailsQ6_K20.89 GB+11.1 GB22–29Derived
Codestral 22B22BQ4_K_M15.03 GB+17 GB32–42Derived
ERNIE 4.5 21B-A3B21B-A3BPair fit detailsQ8_024.48 GB+7.5 GB≤201Ceiling
GPT-OSS 20B21B MoEMXFP412.76 GB+19.2 GB≤307Ceiling

10–20B · 19

Arc Pro B70 32G — what LLMs can it run? — 10–20B
ModelQuantEst. VRAMHeadroomtok/s
InternLM2 20B Chat20BQ5_K_M15.59 GB+16.4 GB30–40Derived
DeepSeek-Coder-V2-Lite Instruct16B-A2.4BQ8_017.56 GB+14.4 GB≤251Ceiling
DeepSeek-V2-Lite Chat16B-A2.4BQ4_K_M10.08 GB+21.9 GB≤440Ceiling
StarCoder2 15B15BQ4_K_M9.85 GB+22.2 GB47–62Derived
Qwen3 14B Instruct14BQ5_K_M11.65 GB+20.4 GB40–54Derived
Qwen2.5 14B Instruct14BQ5_K_M11.73 GB+20.3 GB41–54Derived
DeepSeek-R1-Distill-Qwen-14B14BQ4_K_M10.14 GB+21.9 GB48–63Derived
Phi-4 14B14BQ5_K_M11.77 GB+20.2 GB41–54Derived
Phi-3 Medium 14B Instruct14BQ6_K12.86 GB+19.1 GB37–49Derived
Ministral 3 14B Instruct14BPair fit detailsQ8_015.7 GB+16.3 GB30–39Derived
Mistral Nemo 12B Instruct12BQ6_K11.14 GB+20.9 GB42–56Derived
Gemma 3 12B IT12BQ5_K_M9.84 GB+22.2 GB49–65Derived
Mellum2.1 12B-A2.5B Thinking12B-A2.5BQ8_013.6 GB+18.4 GB≤241Ceiling
Stable LM 2 12B Chat12BQ4_K_M8.52 GB+23.5 GB58–77Derived
Gemma 4 12B IT12BPair fit detailsQ8_013.82 GB+18.2 GB33–44Derived
Llama 3.2 11B Vision Instruct11BQ8_012.9 GB+19.1 GB36–48Derived
Solar 10.7B Instruct11BQ4_K_M7.6 GB+24.4 GB65–87Derived
Falcon 3 10B Instruct10BQ4_K_M7.28 GB+24.7 GB67–89Derived
Metis-1.6 Logos10B MoEPair fit detailsQ8_011.67 GB+20.3 GB≤230Ceiling

4–10B · 37

Arc Pro B70 32G — what LLMs can it run? — 4–10B
ModelQuantEst. VRAMHeadroomtok/s
Gemma 2 9B Instruct9BQ8_011.7 GB+20.3 GB43–57Derived
Qwen3.5 9B9BPair fit detailsQ8_010.39 GB+21.6 GB43–57Derived
K2 Horizon 7B7BPair fit detailsQ8_010.62 GB+21.4 GB44–59Derived
GLM-4-9B-Chat9BQ8_010.16 GB+21.8 GB44–59Derived
MiMo-V2.6-Distill-Qwen-9B9BPair fit detailsQ8_010.11 GB+21.9 GB44–59Derived
Clairic Flash 9B9BPair fit detailsQ8_010.1 GB+21.9 GB44–59Derived
Qwen3-VL 8B Instruct8BQ8_010.39 GB+21.6 GB45–60Derived
TwIL 8B balanced8.8BQ4_K_M6.28 GB+25.7 GB79–105Derived
Granite 4.2 8B8BPair fit detailsQ8_010.46 GB+21.5 GB45–60Derived
Ministral 3 8B Instruct8BQ8_010.02 GB+22 GB47–62Derived
LFM2.5 8B-A1B8B-A1.5BPair fit detailsQ8_09.46 GB+22.5 GB≤401Ceiling
Qwen2-VL 7B Instruct7BQ4_K_M5.49 GB+26.5 GB84–112Derived
Granite 3.1 8B Instruct8BQ4_K_M5.88 GB+26.1 GB85–113Derived
Shot Reader 8B8BPair fit detailsQ8_09.72 GB+22.3 GB49–64Derived
Qwen3 8B Instruct8BQ6_K7.64 GB+24.4 GB63–84Derived
Llama 3.1 8B Instruct8BQ8_09.47 GB+22.5 GB50–66Derived
Nous Hermes 3 Llama 3.1 8B8BQ4_K_M5.64 GB+26.4 GB87–115Derived
Aya 23 8B8BQ4_K_M5.64 GB+26.4 GB87–115Derived
OpenChat 3.6 8B8BQ4_K_M5.64 GB+26.4 GB87–115Derived
DeepSeek-R1-Distill-Llama-8B8BQ5_K_M6.51 GB+25.5 GB74–98Derived
InternLM2 7B Chat7BQ4_K_M5.45 GB+26.6 GB90–120Derived
Qwen2.5 7B Instruct7BQ6_K6.77 GB+25.2 GB68–90Derived
Qwen2.5-Coder 7B Instruct7BQ4_K_M5.07 GB+26.9 GB92–121Derived
DeepSeek-R1-Distill-Qwen-7B7BQ4_K_M5.07 GB+26.9 GB92–121Derived
Gemma 4 E4B ITE4BPair fit detailsQ8_08.68 GB+23.3 GB52–68Derived
Nemotron 3.5 Bolt E4BE4BPair fit detailsQ8_08.5 GB+23.5 GB53–70Derived
OLMo 2 7B Instruct7BQ8_010.3 GB+21.7 GB55–72Derived
OLMo 3 7B Instruct7BQ8_010.3 GB+21.7 GB55–72Derived
WizardLM-2 7B7BQ4_K_M5.14 GB+26.9 GB97–128Derived
Mistral 7B Instruct v0.37BQ6_K6.75 GB+25.3 GB71–95Derived
Zephyr 7B Beta7BQ6_K6.75 GB+25.3 GB71–95Derived
K2 Horizon 3.7B3.7BQ8_06.24 GB+25.8 GB≤119Ceiling
Gemma 4 E2B ITE2BQ8_05.2 GB+26.8 GB≤130Ceiling
Qwen3.5 4B4BQ8_04.97 GB+27 GB≤138Ceiling
Gemma 3 4B IT4BQ8_05.05 GB+27 GB≤140Ceiling
Spark-X2.6 4B4BQ8_04.84 GB+27.2 GB≤146Ceiling
Qwen3 4B Instruct4BQ6_K4.05 GB+28 GB≤195Ceiling

<4B · 22

Arc Pro B70 32G — what LLMs can it run? — <4B
ModelQuantEst. VRAMHeadroomtok/s
Phi-4 Mini Instruct3.8BQ8_04.81 GB+27.2 GB≤157Ceiling
Phi-3.5 Mini Instruct3.8BQ8_05.89 GB+26.1 GB≤158Ceiling
Granite 4.2 3B3BQ8_04.41 GB+27.6 GB≤165Ceiling
Ministral 3 3B Instruct3BQ8_04.26 GB+27.7 GB≤175Ceiling
Llama 3.2 3B Instruct3BQ8_04.05 GB+28 GB≤188Ceiling
Qwen2.5 3B Instruct3BQ8_03.59 GB+28.4 GB≤195Ceiling
SmolLM3 3B3BQ8_03.73 GB+28.3 GB≤195Ceiling
Metis-1.6 Praxis3B MoEQ8_03.4 GB+28.6 GB≤538Ceiling
Gemma 2 2B Instruct2BQ8_03.34 GB+28.7 GB≤231Ceiling
Qwen3.5 2B2BQ8_02.23 GB+29.8 GB≤307Ceiling
LFM3 1.7B1.7BQ8_01.99 GB+30 GB≤347Ceiling
LFM3-VL 2B2BQ8_01.99 GB+30 GB≤347Ceiling
Qwen3 1.7B Instruct1.7BQ8_02.39 GB+29.6 GB≤349Ceiling
Qwen2.5 1.5B Instruct1.5BQ8_01.83 GB+30.2 GB≤392Ceiling
Llama 3.2 1B Instruct1BQ8_01.51 GB+30.5 GB≤486Ceiling
LFM2.5 1.2B Instruct1.2BQ8_01.35 GB+30.7 GB≤515Ceiling
K2 Horizon 0.9B0.9BQ8_01.44 GB+30.6 GB≤558Ceiling
Qwen3.5 0.8B0.8BQ8_00.94 GB+31.1 GB≤751Ceiling
BotPsy 0.8B0.75BQ8_00.9 GB+31.1 GB≤790Ceiling
Qwen2.5 0.5B Instruct0.5BQ8_00.6 GB+31.4 GB≤1241Ceiling
MNN-Pocket 0.3B Instruct0.36BQ8_00.46 GB+31.5 GB≤1643Ceiling
MNN-Pocket 0.3B QAT0.30BQ4_10.26 GB+31.7 GB≤3378Ceiling

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 each estimate leaves headroom. Raise the context length and the list shortens; the calculator lets you check any combination directly. Speed figures are labelled as measurements, estimates scaled from this card's Llama 2 7B benchmark, or theoretical bandwidth ceilings. A figure that only exists for the RTX 4090 is labelled as that card.

A further 3 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.

Measured on this card

No benchmark runs in this index were recorded on a Arc Pro B70 32G. 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 32GB 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: 32 GB RAM (CPU)

    Setup guides for this card

    Step-by-step guides written for the Arc Pro B70 32G's runtime or memory size, with the commands and the checks that prove the model is running on the card.

    Common questions

    What is the best local LLM for an Arc Pro B70 32G?

    For everyday use, Gemma 3 27B IT at Q4_K_M — about 18.3 GB of the card's 32 GB at 4K context, leaving 13.7 GB for a longer window. If you want the largest thing that will load, that is Kolibri-1 78B-A3.5B at IQ2_XS (25.6 GB, 6.4 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 Pro B70 32G run Llama 3.2 90B Vision Instruct?

    No. Its smallest build here, Q3_K_M, needs about 46.2 GB and this card has 32.0 GB usable — short by 14.2 GB before any context beyond 4K. The largest model this card does clear is Kolibri-1 78B-A3.5B.

    How many tokens per second does an Arc Pro B70 32G 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 608 GB/s — a ceiling near 132 tok/s. Batch size, context length, the runtime and how much of the model sits in cache all take you below it.

    Arc Pro B70 32G or RTX 5090 for local LLMs?

    They do not hold the same set: Arc Pro B70 32G fits 123 and RTX 5090 125 of 151 at 4K — the cards load different formats. Speed is the other difference — 608 GB/s against 1,792 GB/s, a 2.9× gap in how fast the weights can be read, which is what token generation is bound by. The RTX 5090 generates faster on any model both can hold.

    123 of 151 indexed models fit comfortably in 32GB 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 2026-11-22 RSS → /feed.xml