Intel · workstation

Arc Pro A30M 4GB

Arc Pro A30M 4GB has 4 GB of VRAM at 112 GB/s — about 3.72 GiB usable after driver and compositor overhead. 508 of 2118 indexed models fit at 32K context with q8_0 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
4 GB
GDDR6
Bandwidth
112 GB/s
64-bit bus
Tensor FP16
—
dense
TDP
50 W
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 388vision language 45audio tts 18audio asr 35embedding 20video 2

What fits at 32K context

largest quantization that fits, per model · 508 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Gemma-3-4b-it-Uncensored-DBL-XI1-Q4_04.7B2.41 GiB0.50 GiB3.72 GiB0.00 GiB21±30%
Vikhr-Gemma-2B-instructQ5_K_L2.6B1.92 GiB0.98 GiB3.72 GiB0.00 GiB21±30%
gemma-2-2b-it-abliteratedQ5_K_L2.6B1.92 GiB0.98 GiB3.72 GiB0.00 GiB21±30%
Gemmasutra-Mini-2B-v1Q5_K_L2.6B1.92 GiB0.98 GiB3.72 GiB0.00 GiB21±30%
OpenClaude-1.7B-MergedQ4_K_M1.7B1.07 GiB1.86 GiB3.72 GiB0.00 GiB21±30%
Qwen3.5-4B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKINGI1-Q4_04.5B2.37 GiB0.53 GiB3.72 GiB0.00 GiB21±30%
Qwen3.5-4B-NSFW-ARA-Heretic-LiteroticaI1-Q4_04.2B2.37 GiB0.53 GiB3.72 GiB0.00 GiB21±30%
Qwen3.5-4B-SOMPOA-heresy-v2I1-Q4_04.5B2.37 GiB0.53 GiB3.72 GiB0.00 GiB21±30%
Qwen3.5-4B-SOMPOA-heresyI1-Q4_04.5B2.37 GiB0.53 GiB3.72 GiB0.00 GiB21±30%
Qwen3.5-4B-Safety-ThinkingI1-Q4_04.2B2.37 GiB0.53 GiB3.72 GiB0.00 GiB21±30%
Huihui-Qwen3.5-4B-abliteratedI1-Q4_04.5B2.37 GiB0.53 GiB3.72 GiB0.00 GiB21±30%
Qwen3.5-4B-RpRMax-v1I1-Q4_04.7B2.37 GiB0.53 GiB3.72 GiB0.00 GiB21±30%
Holo-3.1-4B-uncensored-hereticI1-Q4_04.5B2.37 GiB0.53 GiB3.72 GiB0.00 GiB21±30%
GRaPE-2-MiniI1-Q4_04.7B2.37 GiB0.53 GiB3.72 GiB0.00 GiB21±30%
Darkidol-Ballad-4BI1-Q4_04.5B2.37 GiB0.53 GiB3.72 GiB0.00 GiB21±30%
Qwen3.5-DPO-4B-2I1-Q4_04.2B2.37 GiB0.53 GiB3.72 GiB0.00 GiB21±30%
Huihui-Qwen3.5-4B-Claude-4.6-Opus-abliteratedI1-Q4_04.7B2.37 GiB0.53 GiB3.72 GiB0.00 GiB21±30%
Qwopus3.5-4B-v3-hereticI1-Q4_04.5B2.37 GiB0.53 GiB3.72 GiB0.00 GiB21±30%
Aureth-4B-Qwen3.5I1-Q4_04.5B2.37 GiB0.53 GiB3.72 GiB0.00 GiB21±30%
Supertron2-Reranker-2BI1-Q4_12.1B1.06 GiB1.86 GiB3.72 GiB0.00 GiB21±30%
Uni-MuMER-Qwen3-VL-2BI1-Q4_12.1B1.06 GiB1.86 GiB3.72 GiB0.00 GiB21±30%
Qwen3-VL-2B-ThinkingQ4_12.1B1.06 GiB1.86 GiB3.72 GiB0.00 GiB21±30%
Qwen3-VL-Reranker-2BI1-Q4_12.1B1.06 GiB1.86 GiB3.72 GiB0.00 GiB21±30%
Qwen3-VL-2B-InstructQ4_12.1B1.06 GiB1.86 GiB3.72 GiB0.00 GiB21±30%
OpenCaption-2B-VL-SFT-v1.0I1-Q4_12.1B1.06 GiB1.86 GiB3.72 GiB0.00 GiB21±30%
Atomight-V2.5-1.7BI1-Q4_11.7B1.06 GiB1.86 GiB3.72 GiB0.00 GiB21±30%
gaon-1.7b-v2-translateI1-Q4_11.7B1.06 GiB1.86 GiB3.72 GiB0.00 GiB21±30%
gaon-1.7b-v2-instructI1-Q4_11.7B1.06 GiB1.86 GiB3.72 GiB0.00 GiB21±30%
Lightning-1.7BQ4_11.7B1.06 GiB1.86 GiB3.72 GiB0.00 GiB21±30%
DorsetHeatwaveLLM2I1-Q4_11.7B1.06 GiB1.86 GiB3.72 GiB0.00 GiB21±30%
Qwen2.5-VL-7B-InstructUD-IQ1_S8.3B1.93 GiB0.93 GiB3.72 GiB0.00 GiB21±30%
DeepSeek-OCRMoEQ6_K3.3B2.43 GiB0.50 GiB3.71 GiB0.01 GiB38±37%
Qwen3.5-4BQ4_04.7B2.37 GiB0.53 GiB3.71 GiB0.01 GiB21±30%
Qwen3-1.7BQ3_K_L2.0B1.06 GiB1.86 GiB3.71 GiB0.01 GiB21±30%
AMD-OLMo-1B-SFT-DPOQ5_K_M1.2B0.79 GiB2.13 GiB3.71 GiB0.01 GiB21±30%
Fara1.5-4BIQ4_XS4.5B2.37 GiB0.53 GiB3.71 GiB0.01 GiB21±30%
AREX-TurboIQ4_XS4.5B2.37 GiB0.53 GiB3.71 GiB0.01 GiB21±30%
Darwin-4B-ChimeraI1-Q4_K_S4.0B2.22 GiB0.68 GiB3.71 GiB0.01 GiB21±30%
medgemma-4b-itQ4_K_L4.3B2.47 GiB0.42 GiB3.71 GiB0.01 GiB21±30%
amoral-gemma3-4B-v1Q4_K_L4.3B2.47 GiB0.42 GiB3.71 GiB0.01 GiB21±30%
gemma-3-4b-it-abliteratedQ4_K_L4.3B2.47 GiB0.42 GiB3.71 GiB0.01 GiB21±30%
gemma-3-4b-itQ4_K_L4.3B2.47 GiB0.42 GiB3.71 GiB0.01 GiB21±30%
Tini-Cybersec-8B-A1BMoEIQ2_M8.5B2.71 GiB0.20 GiB3.71 GiB0.01 GiB47±37%
gemma-4-E2B-it-ultra-uncensored-hereticI1-Q2_K5.1B2.78 GiB0.13 GiB3.71 GiB0.01 GiB21±30%
Barcenas-E2BI1-Q2_K5.1B2.78 GiB0.13 GiB3.71 GiB0.01 GiB21±30%
gemma-4-E2B-it-Uncensored-MAXI1-Q2_K5.1B2.78 GiB0.13 GiB3.71 GiB0.01 GiB21±30%
Firefly-v4I1-Q2_K5.1B2.78 GiB0.13 GiB3.71 GiB0.01 GiB21±30%
gemma-4-E2B-it-abliteratedI1-Q2_K5.1B2.78 GiB0.13 GiB3.71 GiB0.01 GiB21±30%
gemma-4-E2B-it-uncensoredQ2_K5.1B2.78 GiB0.13 GiB3.71 GiB0.01 GiB21±30%
gemma-4-E2B-it-heretic-araQ2_K5.1B2.78 GiB0.13 GiB3.71 GiB0.01 GiB21±30%
rp-model_E2B_v5.1I1-Q2_K5.1B2.78 GiB0.13 GiB3.71 GiB0.01 GiB21±30%
BartaLens-E2BI1-Q2_K5.1B2.78 GiB0.13 GiB3.71 GiB0.01 GiB21±30%
gemma-4-E2BQ2_K5.1B2.78 GiB0.13 GiB3.71 GiB0.01 GiB21±30%
granite-3.1-3b-a800m-instructMoEQ4_K_M3.3B1.88 GiB1.06 GiB3.71 GiB0.01 GiB22±37%
Holo-3.1-4BI1-Q3_K_M5.2B2.36 GiB0.53 GiB3.70 GiB0.02 GiB21±30%
AfriqueQwen3.5-4BI1-Q3_K_M5.2B2.36 GiB0.53 GiB3.70 GiB0.02 GiB21±30%
TimeOmni-1-4BI1-Q3_K_M5.2B2.36 GiB0.53 GiB3.70 GiB0.02 GiB21±30%
ToriiGate-0.5Q3_K_M5.2B2.36 GiB0.53 GiB3.70 GiB0.02 GiB21±30%
chandra-ocr-2Q3_K_M5.3B2.36 GiB0.53 GiB3.70 GiB0.02 GiB21±30%
gemma-4-E2B-itIQ4_XS5.1B2.78 GiB0.13 GiB3.70 GiB0.02 GiB21±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 A30M 4GB run?
508 of 2118 indexed open-weight models fit a Arc Pro A30M 4GB at 32,768 context with q8_0 KV cache, the largest being Gemma-3-4b-it-Uncensored-DBL-X at I1-Q4_0. That covers text, vision-language, image, video and speech models.
How much usable memory does a Arc Pro A30M 4GB actually have?
Its nameplate is 4 GB, but about 3.72 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Arc Pro A30M 4GB fast for local AI?
Its memory bandwidth is 112 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.
Arc Pro A30M 4GB — what AI models can it run locally? — ossmodeldb