What AI models can a Arc Pro B70 32GB run?

2022 of 2118 indexed open-weight models fit a Arc Pro B70 32GB at 8,192 context with q8_0 KV cache, the largest being Llama-3_3-Nemotron-Super-49B-v1_5 at IQ3_XXS. That covers text, vision-language, image, video and speech models.

How much usable memory does a Arc Pro B70 32GB actually have?

Its nameplate is 32 GB, but about 29.76 GiB is available to a model once driver and compositor overhead is accounted for.

Is a Arc Pro B70 32GB fast for local AI?

Its memory bandwidth is 608 GB/s, and that figure — not teraflops — is what governs token generation speed. Capacity decides what you can run; bandwidth decides how fast it runs.

Intel · workstation

Arc Pro B70 32GB

Arc Pro B70 32GB has 32 GB of VRAM at 608 GB/s — about 29.76 GiB usable after driver and compositor overhead. 2022 of 2118 indexed models fit at 8K context with q8_0 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed

Memory

32 GB

GDDR6

Bandwidth

608 GB/s

256-bit bus

Tensor FP16

184 TF

dense

TDP

230 W

$949 MSRP

Context4K 8K 16K 32K 64K 128K

KV cachef16 q8_0 q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context

text 1737audio tts 21vision language 181video 16image 2embedding 26audio asr 39

What fits at 8K context

largest quantization that fits, per model · 2022 of 2118 indexed

Model	Best quant	Params	Weights●	KV●	Total in memory◐	Headroom◐	tok/s◐
Llama-3_3-Nemotron-Super-49B-v1_5	IQ3_XXS	49.9B	18.18 GiB	10.63 GiB	29.74 GiB	0.02 GiB	12±30%
Valkyrie-49B-v2.1	I1-IQ3_XXS	49.9B	18.18 GiB	10.63 GiB	29.74 GiB	0.02 GiB	12±30%
Llama-3_3-Nemotron-Super-49B-v1	IQ3_XXS	49.9B	18.18 GiB	10.63 GiB	29.74 GiB	0.02 GiB	12±30%
Qwen3-TTS-12Hz-0.6B-Base	F32	915M	28.88 GiB	0.00 GiB	29.72 GiB	0.04 GiB	11±30%
OLMo-2-1124-13B-Instruct	F16	13.7B	25.55 GiB	3.32 GiB	29.72 GiB	0.04 GiB	11±30%
Llama-4-Scout-17B-16E-InstructMoEKV unresolved	IQ2_XXS	109B	28.09 GiB	0.80 GiB	29.72 GiB	0.04 GiB	43±37%
Qwen3.5-35B-A3BMoE	Q6_K	36.0B	28.82 GiB	0.08 GiB	29.71 GiB	0.05 GiB	64±37%
Qwen3.6-35B-A3BMoE	Q6_K	36.0B	28.82 GiB	0.08 GiB	29.71 GiB	0.05 GiB	64±37%
Qwen3-Coder-Next-Opus-4.6-Reasoning-DistilledMoE	IQ3_XXS	—	28.78 GiB	0.10 GiB	29.67 GiB	0.09 GiB	71±37%
Gemma-3-27B-MeditronFO	Q8_0	28.8B	28.13 GiB	0.66 GiB	29.66 GiB	0.10 GiB	12±30%
Hypernova-60B-2605MoE	I1-IQ3_S	58.7B	28.73 GiB	0.15 GiB	29.66 GiB	0.10 GiB	54±37%
Llama-3_1-Nemotron-51B-Instruct	Q2_K	51.5B	18.09 GiB	10.63 GiB	29.65 GiB	0.11 GiB	12±30%
Qwen3-53B-A3B-2507-THINKING-TOTAL-RECALL-v2-MASTER-CODERMoE	I1-Q4_K_S	53.0B	28.13 GiB	0.70 GiB	29.62 GiB	0.14 GiB	40±37%
Seed-OSS-36B-Instruct-biprojected-norm-preserving-abliterated	I1-Q6_K	36.2B	27.63 GiB	1.06 GiB	29.59 GiB	0.17 GiB	12±30%
Seed-OSS-36B-Instruct	Q6_K	36.2B	27.63 GiB	1.06 GiB	29.59 GiB	0.17 GiB	12±30%
Hermes-4.3-36B-heretic	I1-Q6_K	36.2B	27.63 GiB	1.06 GiB	29.59 GiB	0.17 GiB	12±30%
Hermes-4.3-36B	Q6_K	36.2B	27.63 GiB	1.06 GiB	29.59 GiB	0.17 GiB	12±30%
Seed-OSS-36B-Base	Q6_K	36.2B	27.63 GiB	1.06 GiB	29.59 GiB	0.17 GiB	12±30%
Rombo-LLM-V3.0-Qwen-72b	I1-IQ2_M	72.7B	27.32 GiB	1.33 GiB	29.58 GiB	0.18 GiB	12±30%
Qwen2.5-72B-Instruct-abliterated	I1-IQ2_M	72.7B	27.32 GiB	1.33 GiB	29.58 GiB	0.18 GiB	12±30%
Qwen2.5-72B-Instruct-abliterated-v2	I1-IQ2_M	72.7B	27.32 GiB	1.33 GiB	29.58 GiB	0.18 GiB	12±30%
HuatuoGPT-o1-72B	IQ2_M	72.7B	27.32 GiB	1.33 GiB	29.58 GiB	0.18 GiB	12±30%
MiroThinker-v1.0-72B	I1-IQ2_M	72.7B	27.32 GiB	1.33 GiB	29.58 GiB	0.18 GiB	12±30%
EVA-Qwen2.5-72B-v0.2	IQ2_M	72.7B	27.32 GiB	1.33 GiB	29.58 GiB	0.18 GiB	12±30%
Qwen2.5-Math-72B-Instruct	IQ2_M	72.7B	27.32 GiB	1.33 GiB	29.58 GiB	0.18 GiB	12±30%
Qwen2.5-72B-Instruct	IQ2_M	72.7B	27.32 GiB	1.33 GiB	29.58 GiB	0.18 GiB	12±30%
Malaysian-Qwen2.5-72B-Instruct	I1-IQ2_M	72.7B	27.32 GiB	1.33 GiB	29.58 GiB	0.18 GiB	12±30%
Qwen2.5-72B	I1-IQ2_M	72.7B	27.32 GiB	1.33 GiB	29.58 GiB	0.18 GiB	12±30%
magnum-v4-72b	I1-IQ2_M	72.7B	27.32 GiB	1.33 GiB	29.58 GiB	0.18 GiB	12±30%
KAT-Dev-72B-Exp	IQ2_M	72.7B	27.32 GiB	1.33 GiB	29.58 GiB	0.18 GiB	12±30%
Homer-v1.0-Qwen2.5-72B	IQ2_M	72.7B	27.32 GiB	1.33 GiB	29.58 GiB	0.18 GiB	12±30%
Tower-Plus-72B-ultra-uncensored-heretic	I1-IQ2_M	72.7B	27.32 GiB	1.33 GiB	29.58 GiB	0.18 GiB	12±30%
Qwen2.5-VL-72B-Instruct	IQ2_M	73.4B	27.32 GiB	1.33 GiB	29.58 GiB	0.18 GiB	12±30%
Chronos-Platinum-72B	IQ2_M	72.7B	27.32 GiB	1.33 GiB	29.58 GiB	0.18 GiB	12±30%
UI-TARS-72B-DPO	IQ2_M	73.4B	27.32 GiB	1.33 GiB	29.58 GiB	0.18 GiB	12±30%
Snowpiercer-15B-v4	BF16	15.0B	27.90 GiB	0.83 GiB	29.58 GiB	0.18 GiB	12±30%
Huihui-Qwen3-Coder-Next-abliteratedMoE	I1-IQ3_XXS	79.7B	28.68 GiB	0.10 GiB	29.57 GiB	0.19 GiB	71±37%
Bernini-R	Q8_0	14.3B	28.71 GiB	0.00 GiB	29.56 GiB	0.20 GiB	12±30%
Maenad-70B	I1-IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-Reasoner	I1-IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%
Rombos-LLM-70b-Llama-3.3	I1-IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%
L3.3-Electra-R1-70b	I1-IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%
Latxa-Llama-3.1-70B-Instruct-v2	I1-IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%
Llama-3.3_70_b_uncensored_continued	I1-IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%
Llama-3.3-70B-Instruct-abliterated	I1-IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%
grok-oss-Revenant-70B	I1-IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%
Llama-3.1-Nemotron-70B-Instruct-HF	I1-IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%
L3.3-70B-Euryale-v2.3	I1-IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%
Hermes-3-Llama-3.1-70B	IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%
Hermes-4-70B-heretic	I1-IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%
Llama-3.3-70B-Instruct	IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%
Llama-3.1-70B	IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%
Anubis-70B-v1.2	IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%
Hermes-4-70B	IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%
Golem-70B-v1b	I1-IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%
DeepSeek-R1-Distill-Llama-70B-abliterated	I1-IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%
DeepSeek-R1-Distill-Llama-70B-heretic	I1-IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%
llama-3-firefunction-v2	IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%
Legion-V2.1-LLaMa-70B	I1-IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%
Assistant_Pepe_70B	I1-IQ3_XS	70.6B	27.29 GiB	1.33 GiB	29.55 GiB	0.21 GiB	12±30%

●From the file◐Predictedwhat these mean

Speed is modeled, not measured: decode is memory-bandwidth bound, so tokens per second is bytes read per token against achievable bandwidth. Mixture-of-experts models carry a wider band because only the routed experts are read each step, and few have been measured publicly.

Questions people ask

What AI models can a Arc Pro B70 32GB run?: 2022 of 2118 indexed open-weight models fit a Arc Pro B70 32GB at 8,192 context with q8_0 KV cache, the largest being Llama-3_3-Nemotron-Super-49B-v1_5 at IQ3_XXS. That covers text, vision-language, image, video and speech models.
How much usable memory does a Arc Pro B70 32GB actually have?: Its nameplate is 32 GB, but about 29.76 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Arc Pro B70 32GB fast for local AI?: Its memory bandwidth is 608 GB/s, and that figure — not teraflops — is what governs token generation speed. Capacity decides what you can run; bandwidth decides how fast it runs.