Intel · consumer

Arc A730M 12GB

Arc A730M 12GB has 12 GB of VRAM at 336 GB/s — about 11.16 GiB usable after driver and compositor overhead. 977 of 2118 indexed models fit at 64K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
12 GB
GDDR6
Bandwidth
336 GB/s
192-bit bus
Tensor FP16
—
dense
TDP
120 W
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 784video 14embedding 22audio asr 38audio tts 20vision language 98image 1

What fits at 64K context

largest quantization that fits, per model · 977 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
OLMoE-1B-7B-0924-InstructMoEI1-Q2_K6.9B2.39 GiB8.00 GiB11.16 GiB0.00 GiB12±37%
Wan2.1-FLF2V-14B-720PQ4_116.4B10.32 GiB0.00 GiB11.16 GiB0.00 GiB18±30%
Wan2.1-I2V-14B-480PQ4_116.4B10.32 GiB0.00 GiB11.15 GiB0.01 GiB18±30%
Wan2.1-I2V-14B-720PQ4_116.4B10.32 GiB0.00 GiB11.15 GiB0.01 GiB18±30%
Nemotron-3-Embed-8B-BF16IQ1_S8.0B1.81 GiB8.50 GiB11.15 GiB0.01 GiB18±30%
granite-speech-4.1-2b-narF162.3B5.36 GiB5.00 GiB11.15 GiB0.01 GiB17±30%
Ace-Step1.5Q4_K160M7.32 GiB3.04 GiB11.15 GiB0.01 GiB17±30%
Muse-Glimmer-30BIQ2_S29.8B9.35 GiB0.91 GiB11.14 GiB0.02 GiB18±30%
Phi-4-mini-instruct-abliteratedQ4_K_M3.8B2.32 GiB8.00 GiB11.13 GiB0.03 GiB18±30%
Phi-4-mini-reasoningQ4_K_M3.8B2.32 GiB8.00 GiB11.13 GiB0.03 GiB18±30%
Phi-4-mini-instructQ4_K_M3.8B2.32 GiB8.00 GiB11.13 GiB0.03 GiB18±30%
Qwen3-VL-4B-Instruct-Unredacted-MAXI1-IQ2_S4.4B1.32 GiB9.00 GiB11.13 GiB0.03 GiB18±30%
Qwen3-VL-4B-Thinking-Unredacted-MAXI1-IQ2_S4.4B1.32 GiB9.00 GiB11.13 GiB0.03 GiB18±30%
Zubr1.0-VL-4BI1-IQ2_S4.4B1.32 GiB9.00 GiB11.13 GiB0.03 GiB18±30%
Huihui-Qwen3-VL-4B-Instruct-abliteratedI1-IQ2_S4.4B1.32 GiB9.00 GiB11.13 GiB0.03 GiB18±30%
Qwen3-VL-4B-Instruct-UncensoredI1-IQ2_S4.4B1.32 GiB9.00 GiB11.13 GiB0.03 GiB18±30%
OpenCaption-4B-VL-SFT-v1.0I1-IQ2_S4.4B1.32 GiB9.00 GiB11.13 GiB0.03 GiB18±30%
Parable-Qwen3-4B-Claude-Fable-5I1-IQ2_S4.0B1.32 GiB9.00 GiB11.13 GiB0.03 GiB18±30%
Qwen3-4b-Z-Image-Turbo-AbliteratedV1I1-IQ2_S4.0B1.32 GiB9.00 GiB11.13 GiB0.03 GiB18±30%
Neuron-4B-InstructI1-IQ2_S4.0B1.32 GiB9.00 GiB11.13 GiB0.03 GiB18±30%
ChineseErrorCorrector4-4BI1-IQ2_S4.0B1.32 GiB9.00 GiB11.13 GiB0.03 GiB18±30%
FastContext-1.0-4B-SFT-abliteratedI1-IQ2_S4.0B1.32 GiB9.00 GiB11.13 GiB0.03 GiB18±30%
Qwen3-4B-Instruct_NSFW-V2.1I1-IQ2_S4.0B1.32 GiB9.00 GiB11.13 GiB0.03 GiB18±30%
FastContext-1.0-4B-SFTI1-IQ2_S4.0B1.32 GiB9.00 GiB11.13 GiB0.03 GiB18±30%
fable-traces-abliteratedI1-IQ2_S4.0B1.32 GiB9.00 GiB11.13 GiB0.03 GiB18±30%
Nexa-AI-4B-InstructI1-IQ2_S4.0B1.32 GiB9.00 GiB11.13 GiB0.03 GiB18±30%
Lumen-4B-InstructI1-IQ2_S4.0B1.32 GiB9.00 GiB11.13 GiB0.03 GiB18±30%
Qwen3-HereticLM-4BI1-IQ2_S4.0B1.32 GiB9.00 GiB11.13 GiB0.03 GiB18±30%
Marco-Nano-InstructMoEI1-IQ3_XXS8.0B3.35 GiB7.00 GiB11.12 GiB0.04 GiB14±37%
legitus-instruct-v1I1-IQ2_XXS8.1B2.25 GiB8.00 GiB11.12 GiB0.04 GiB18±30%
Apertus-8B-Instruct-2509I1-IQ2_XXS8.1B2.25 GiB8.00 GiB11.12 GiB0.04 GiB18±30%
Nemotron-Mini-4B-InstructIQ4_XS4.2B2.29 GiB8.00 GiB11.11 GiB0.05 GiB18±30%
Hubble-4B-v1Q3_K_L4.5B2.30 GiB8.00 GiB11.11 GiB0.05 GiB18±30%
Aura-4BI1-Q3_K_L4.5B2.30 GiB8.00 GiB11.11 GiB0.05 GiB18±30%
magnum-v2-4bI1-Q3_K_L4.5B2.30 GiB8.00 GiB11.11 GiB0.05 GiB18±30%
Impish_LLAMA_4BQ3_K_L4.5B2.30 GiB8.00 GiB11.11 GiB0.05 GiB18±30%
Llama-3.1-Minitron-4B-Width-BaseQ3_K_L4.5B2.30 GiB8.00 GiB11.11 GiB0.05 GiB18±30%
Wan2.2-Distill-ModelsQ5_114.3B10.27 GiB0.00 GiB11.10 GiB0.06 GiB18±30%
Marco-Mini-InstructMoEI1-IQ1_S17.3B3.33 GiB7.00 GiB11.10 GiB0.06 GiB14±37%
Hunyuan-7B-InstructIQ2_XS7.5B2.26 GiB8.00 GiB11.10 GiB0.06 GiB18±30%
Bernini-RQ5_114.3B10.26 GiB0.00 GiB11.10 GiB0.06 GiB18±30%
SkyReels-V2-DF-14B-540PQ5_114.3B10.27 GiB0.00 GiB11.10 GiB0.06 GiB18±30%
CycleGRPO-4BI1-IQ2_XXS4.8B1.28 GiB9.00 GiB11.09 GiB0.07 GiB18±30%
Voxtral-Mini-3B-2507Q4_K_M4.7B2.78 GiB7.50 GiB11.09 GiB0.07 GiB18±30%
InternVL3_5-14BQ5_K_L15.1B10.24 GiB0.00 GiB11.08 GiB0.08 GiB18±30%
orpheus-3b-0.1-ftQ8_03.8B3.27 GiB7.00 GiB11.08 GiB0.08 GiB18±30%
Ministral-8B-Instruct-2410Q4_K_M8.0B4.57 GiB5.68 GiB11.08 GiB0.08 GiB18±30%
Foundation-Sec-8B-InstructI1-IQ2_XXS8.0B2.23 GiB8.00 GiB11.07 GiB0.09 GiB18±30%
Foundation-Sec-8B-Instruct-hereticI1-IQ2_XXS8.0B2.23 GiB8.00 GiB11.07 GiB0.09 GiB18±30%
Meta-Llama-3-8BIQ2_XXS8.0B2.23 GiB8.00 GiB11.07 GiB0.09 GiB18±30%
llama3.1-heretic-unsensoredI1-IQ2_XXS8.0B2.23 GiB8.00 GiB11.07 GiB0.09 GiB18±30%
Anubis-Mini-8B-v1I1-IQ2_XXS8.0B2.23 GiB8.00 GiB11.07 GiB0.09 GiB18±30%
dolphin-2.9-llama3-8bIQ2_XXS8.0B2.23 GiB8.00 GiB11.07 GiB0.09 GiB18±30%
Meta-Llama-3-8BIQ2_XXS8.0B2.23 GiB8.00 GiB11.07 GiB0.09 GiB18±30%
L3.1-Dark-Reasoning-LewdPlay-evo-Hermes-R1-Uncensored-8B-hereticI1-IQ2_XXS8.0B2.23 GiB8.00 GiB11.07 GiB0.09 GiB18±30%
L3.1-Dark-Reasoning-LewdPlay-evo-Hermes-R1-Uncensored-8BI1-IQ2_XXS8.0B2.23 GiB8.00 GiB11.07 GiB0.09 GiB18±30%
Gluon-8BI1-IQ2_XXS8.0B2.23 GiB8.00 GiB11.07 GiB0.09 GiB18±30%
Llama3.3-8B-Instruct-Thinking-Heretic-Uncensored-Claude-4.5-Opus-High-ReasoningI1-IQ2_XXS8.0B2.23 GiB8.00 GiB11.07 GiB0.09 GiB18±30%
Llama3.3-8B-Instruct-Thinking-Claude-4.5-Opus-High-ReasoningI1-IQ2_XXS8.0B2.23 GiB8.00 GiB11.07 GiB0.09 GiB18±30%
Hypnos-i1-8BI1-IQ2_XXS8.0B2.23 GiB8.00 GiB11.07 GiB0.09 GiB18±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 A730M 12GB run?
977 of 2118 indexed open-weight models fit a Arc A730M 12GB at 65,536 context with f16 KV cache, the largest being OLMoE-1B-7B-0924-Instruct at I1-Q2_K. That covers text, vision-language, image, video and speech models.
How much usable memory does a Arc A730M 12GB actually have?
Its nameplate is 12 GB, but about 11.16 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Arc A730M 12GB fast for local AI?
Its memory bandwidth is 336 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.