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. 2012 of 2118 indexed models fit at 16K 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
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1727audio tts 21vision language 181video 16image 2embedding 26audio asr 39

What fits at 16K context

largest quantization that fits, per model · 2012 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
c4ai-command-r-plus-08-2024IQ2_XXS104B26.65 GiB2.13 GiB29.75 GiB0.01 GiB12±30%
GLM-4.5-Air-REAP-82B-A12BMoEIQ1_M81.9B27.38 GiB1.53 GiB29.73 GiB0.03 GiB34±37%
Qwen3-TTS-12Hz-0.6B-BaseF32915M28.88 GiB0.00 GiB29.72 GiB0.04 GiB11±30%
NousCoder-14BBF1614.8B27.51 GiB1.33 GiB29.70 GiB0.06 GiB12±30%
Qwen3-14B-Claude-4.5-Opus-High-Reasoning-DistillBF1614.8B27.51 GiB1.33 GiB29.70 GiB0.06 GiB12±30%
Qwen3-14BBF1614.8B27.51 GiB1.33 GiB29.70 GiB0.06 GiB12±30%
Qwen3-14B-abliteratedBF1614.8B27.51 GiB1.33 GiB29.70 GiB0.06 GiB12±30%
Josiefied-Qwen3-14B-abliterated-v3BF1614.8B27.51 GiB1.33 GiB29.70 GiB0.06 GiB12±30%
Hermes-4-14BBF1614.8B27.51 GiB1.33 GiB29.70 GiB0.06 GiB12±30%
Qwen3-14B-BaseF1614.8B27.51 GiB1.33 GiB29.70 GiB0.06 GiB12±30%
Qwen2.5-7B-Instruct-1MF327.6B28.38 GiB0.46 GiB29.69 GiB0.07 GiB12±30%
DeepSeek-R1-Distill-Qwen-7BF327.6B28.38 GiB0.46 GiB29.69 GiB0.07 GiB12±30%
UI-TARS-7B-DPOF328.3B28.38 GiB0.46 GiB29.69 GiB0.07 GiB12±30%
Qwen2-7B-InstructF327.6B28.38 GiB0.46 GiB29.69 GiB0.07 GiB12±30%
Hercules-5.0-Qwen2-7BF327.6B28.38 GiB0.46 GiB29.69 GiB0.07 GiB12±30%
Kepler-8B-Instruct-v2F167.6B28.37 GiB0.46 GiB29.69 GiB0.07 GiB12±30%
MiniCPM-o-2_6F328.7B28.37 GiB0.46 GiB29.69 GiB0.07 GiB12±30%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedUD-TQ1_0109B27.25 GiB1.59 GiB29.67 GiB0.09 GiB37±37%
Huihui-Qwen3-Coder-Next-abliteratedMoEI1-IQ3_XXS79.7B28.68 GiB0.20 GiB29.67 GiB0.09 GiB68±37%
Yi-34B-200K-DARE-megamerge-v8Q6_K34.4B26.78 GiB1.99 GiB29.66 GiB0.10 GiB12±30%
Nous-Hermes-2-Yi-34BI1-Q6_K34.4B26.78 GiB1.99 GiB29.66 GiB0.10 GiB12±30%
Nous-Capybara-limarpv3-34BI1-Q6_K34.4B26.78 GiB1.99 GiB29.66 GiB0.10 GiB12±30%
Apertus-70B-Instruct-2509UD-IQ3_XXS70.6B26.01 GiB2.66 GiB29.65 GiB0.11 GiB12±30%
Huihui-GLM-4.7-Flash-abliterated-57BMoEI1-IQ4_XS57.3B27.68 GiB1.11 GiB29.63 GiB0.13 GiB40±37%
Salience-1.5-ProMoEQ6_K_L36.0B28.66 GiB0.17 GiB29.63 GiB0.13 GiB62±37%
Qwable-v1MoEQ6_K_L36.0B28.66 GiB0.17 GiB29.63 GiB0.13 GiB62±37%
T-SearchMoEQ6_K_L36.0B28.66 GiB0.17 GiB29.63 GiB0.13 GiB62±37%
Phi-3.5-MoE-instructMoEKV unresolvedQ5_K_L41.9B27.75 GiB1.06 GiB29.62 GiB0.14 GiB30±37%
Rombo-LLM-V3.0-Qwen-72bI1-IQ2_S72.7B26.02 GiB2.66 GiB29.61 GiB0.15 GiB12±30%
Qwen2.5-72B-Instruct-abliteratedI1-IQ2_S72.7B26.02 GiB2.66 GiB29.61 GiB0.15 GiB12±30%
Qwen2.5-72B-Instruct-abliterated-v2I1-IQ2_S72.7B26.02 GiB2.66 GiB29.61 GiB0.15 GiB12±30%
HuatuoGPT-o1-72BIQ2_S72.7B26.02 GiB2.66 GiB29.61 GiB0.15 GiB12±30%
MiroThinker-v1.0-72BI1-IQ2_S72.7B26.02 GiB2.66 GiB29.61 GiB0.15 GiB12±30%
Malaysian-Qwen2.5-72B-InstructI1-IQ2_S72.7B26.02 GiB2.66 GiB29.61 GiB0.15 GiB12±30%
Qwen2.5-72BI1-IQ2_S72.7B26.02 GiB2.66 GiB29.61 GiB0.15 GiB12±30%
magnum-v4-72bI1-IQ2_S72.7B26.02 GiB2.66 GiB29.61 GiB0.15 GiB12±30%
KAT-Dev-72B-ExpIQ2_S72.7B26.02 GiB2.66 GiB29.61 GiB0.15 GiB12±30%
Homer-v1.0-Qwen2.5-72BIQ2_S72.7B26.02 GiB2.66 GiB29.61 GiB0.15 GiB12±30%
Tower-Plus-72B-ultra-uncensored-hereticI1-IQ2_S72.7B26.02 GiB2.66 GiB29.61 GiB0.15 GiB12±30%
Qwen2.5-VL-72B-InstructIQ2_S73.4B26.02 GiB2.66 GiB29.61 GiB0.15 GiB12±30%
UI-TARS-72B-DPOIQ2_S73.4B26.02 GiB2.66 GiB29.61 GiB0.15 GiB12±30%
HarmonicHarlequin_v5-20BI1-Q2_K33.3B11.47 GiB17.27 GiB29.58 GiB0.18 GiB12±30%
Bernini-RQ8_014.3B28.71 GiB0.00 GiB29.56 GiB0.20 GiB12±30%
command-r-35b-writer-v2IQ4_XS35.0B18.02 GiB10.63 GiB29.55 GiB0.21 GiB12±30%
Gemma-4-Novelist-Eclipse-31BQ6_K_L32.7B26.60 GiB1.95 GiB29.44 GiB0.32 GiB12±30%
Gemma-4-31B-StyleTuneQ6_K_L32.7B26.60 GiB1.95 GiB29.44 GiB0.32 GiB12±30%
MiniMax-M2.1-REAP-139B-A10BMoEI1-IQ1_S139B26.55 GiB2.06 GiB29.39 GiB0.37 GiB36±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-IQ1_S139B26.55 GiB2.06 GiB29.39 GiB0.37 GiB36±37%
Assistant_Pepe_70BQ2_K70.6B25.79 GiB2.66 GiB29.38 GiB0.38 GiB12±30%
Hermes-4-70BUD-IQ3_XXS70.6B25.76 GiB2.66 GiB29.34 GiB0.42 GiB12±30%
Llama-3.3-70B-InstructUD-IQ3_XXS70.6B25.76 GiB2.66 GiB29.34 GiB0.42 GiB12±30%
DeepSeek-R1-Distill-Llama-70BUD-IQ3_XXS70.6B25.76 GiB2.66 GiB29.34 GiB0.42 GiB12±30%
Apriel-1.6-15b-ThinkerBF1614.9B26.88 GiB1.59 GiB29.33 GiB0.43 GiB12±30%
Qwen3.5-88BMoEI1-Q2_K_S87.7B28.30 GiB0.20 GiB29.32 GiB0.44 GiB56±37%
Kimi-Linear-48B-A3B-InstructMoEQ4_K_L49.1B28.26 GiB0.25 GiB29.32 GiB0.44 GiB12±30%
ALIA-40b-fc-2606I1-Q5_K_M40.4B26.78 GiB1.59 GiB29.29 GiB0.47 GiB12±30%
ALIA-40b-instruct-2606I1-Q5_K_M40.4B26.78 GiB1.59 GiB29.29 GiB0.47 GiB12±30%
Hunyuan-A13B-InstructMoEQ2_K80.4B27.40 GiB1.06 GiB29.26 GiB0.50 GiB12±30%
Qwen3-Coder-NextMoEUD-IQ3_S79.7B27.65 GiB0.80 GiB29.24 GiB0.52 GiB58±37%
CalmeRys-78B-Orpo-v0.1I1-IQ2_XXS78.0B25.43 GiB2.86 GiB29.21 GiB0.55 GiB12±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?
2012 of 2118 indexed open-weight models fit a Arc Pro B70 32GB at 16,384 context with q8_0 KV cache, the largest being c4ai-command-r-plus-08-2024 at IQ2_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.