Intel · consumer

Arc A350M 4GB

Arc A350M 4GB has 4 GB of VRAM at 112 GB/s — about 3.72 GiB usable after driver and compositor overhead. 850 of 2118 indexed models fit at 4K context with q4_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
35 W
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
vision language 59text 707audio asr 37image 1embedding 25video 2audio tts 19

What fits at 4K context

largest quantization that fits, per model · 850 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
gemma-4-E2B-it-qat-q4_0-unquantized-hereticI1-IQ3_M5.1B2.91 GiB0.02 GiB3.72 GiB0.00 GiB21±30%
Huihui-gemma-4-E2B-it-qat-q4_0-unquantized-abliteratedI1-IQ3_M5.1B2.91 GiB0.02 GiB3.72 GiB0.00 GiB21±30%
Gemma4_E2B_Abliterated_Baked_HF_ReadyI1-IQ3_M5.1B2.91 GiB0.02 GiB3.72 GiB0.00 GiB21±30%
Dolphin3.0-Qwen2.5-1.5BF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
Qwen2.5-1.5B-Instruct-abliteratedF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
Qwen2.5-1.5B-VibeThinker-heretic-uncensored-abliteratedF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
NEXUS-Coder-OBLITERATEDF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
NEXUS-Coder-AbliteratedF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
Qwen2.5-1.5B-hereticF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
Qwen2.5-Math-1.5B-InstructBF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
ShellWhisperer-1.5BF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
Qwen2.5-1.5BF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
Qwen2.5-1.5B-InstructF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
Qwen2.5-Coder-1.5B-InstructF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
PiCo-1BF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
Qwen2-VL-2B-InstructF162.2B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
FableForge-1.5BF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
Qwen2-1.5B-InstructBF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
Qwen2-1.5BBF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
Qwen2.5-Coder-1.5BF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
NEXUS-MedicalF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
NEXUS-ScienceF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
NEXUS-LegalF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
NEXUS-CoderF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
NEXUS-FinanceF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
NEXUS-SecurityF161.5B2.88 GiB0.03 GiB3.71 GiB0.01 GiB21±30%
Gemma-3-4b-it-Uncensored-DBL-XI1-Q5_K_S4.7B2.82 GiB0.08 GiB3.71 GiB0.01 GiB21±30%
NVIDIA-Nemotron-3-Nano-4B-BF16Q4_K_S4.0B2.70 GiB0.18 GiB3.71 GiB0.01 GiB21±30%
Agents-A1-4B-Heretic-ARA-Refusals8Q5_K_M4.5B2.86 GiB0.04 GiB3.71 GiB0.01 GiB21±30%
Qwen3.5-4B-NSFW-ARA-Heretic-LiteroticaI1-Q5_K_M4.2B2.86 GiB0.04 GiB3.71 GiB0.01 GiB21±30%
Qwen3.5-4B-RpRMax-v1I1-Q5_K_M4.7B2.86 GiB0.04 GiB3.71 GiB0.01 GiB21±30%
Holo-3.1-4B-uncensored-hereticI1-Q5_K_M4.5B2.86 GiB0.04 GiB3.71 GiB0.01 GiB21±30%
GRaPE-2-MiniI1-Q5_K_M4.7B2.86 GiB0.04 GiB3.71 GiB0.01 GiB21±30%
Qwen3.5-DPO-4B-2I1-Q5_K_M4.2B2.86 GiB0.04 GiB3.71 GiB0.01 GiB21±30%
Qwen3.5-4B-BaseQ5_K_M4.7B2.86 GiB0.04 GiB3.71 GiB0.01 GiB21±30%
Qwen3.5-4B-hereticQ5_K_M4.5B2.86 GiB0.04 GiB3.71 GiB0.01 GiB21±30%
NuExtract3Q5_K_M4.5B2.86 GiB0.04 GiB3.71 GiB0.01 GiB21±30%
Huihui-Qwen3.5-4B-Claude-4.6-Opus-abliteratedI1-Q5_K_M4.7B2.86 GiB0.04 GiB3.71 GiB0.01 GiB21±30%
Qwopus3.5-4B-v3-hereticI1-Q5_K_M4.5B2.86 GiB0.04 GiB3.71 GiB0.01 GiB21±30%
Aureth-4B-Qwen3.5I1-Q5_K_M4.5B2.86 GiB0.04 GiB3.71 GiB0.01 GiB21±30%
FrickFritz-4BI1-Q5_K_S4.7B2.86 GiB0.04 GiB3.71 GiB0.01 GiB21±30%
qwen3.5-4b-agentic-coder-v4I1-Q5_K_S4.7B2.86 GiB0.04 GiB3.71 GiB0.01 GiB21±30%
Newton-bot-3-VLM-mini-4BQ5_K_S4.7B2.86 GiB0.04 GiB3.71 GiB0.01 GiB21±30%
Myth-4BI1-Q5_K_S4.3B2.86 GiB0.04 GiB3.71 GiB0.01 GiB21±30%
Qwen3.5-4B-UncensoredI1-Q5_K_S4.7B2.86 GiB0.04 GiB3.71 GiB0.01 GiB21±30%
JOSIE-2-4B-PreviewI1-Q5_K_S4.7B2.86 GiB0.04 GiB3.71 GiB0.01 GiB21±30%
Surogate-3.5-4BI1-Q5_K_S5.3B2.86 GiB0.04 GiB3.71 GiB0.01 GiB21±30%
Qwopus3.5-4B-v3Q5_K_S4.7B2.86 GiB0.04 GiB3.71 GiB0.01 GiB21±30%
gemma-4-E2B-it-ultra-uncensored-hereticI1-IQ3_S5.1B2.90 GiB0.02 GiB3.71 GiB0.01 GiB21±30%
Barcenas-E2BI1-IQ3_S5.1B2.90 GiB0.02 GiB3.71 GiB0.01 GiB21±30%
gemma-4-E2B-it-Uncensored-MAXI1-IQ3_S5.1B2.90 GiB0.02 GiB3.71 GiB0.01 GiB21±30%
Firefly-v4I1-IQ3_S5.1B2.90 GiB0.02 GiB3.71 GiB0.01 GiB21±30%
gemma-4-E2B-it-abliteratedI1-IQ3_S5.1B2.90 GiB0.02 GiB3.71 GiB0.01 GiB21±30%
rp-model_E2B_v5.1I1-IQ3_S5.1B2.90 GiB0.02 GiB3.71 GiB0.01 GiB21±30%
BartaLens-E2BI1-IQ3_S5.1B2.90 GiB0.02 GiB3.71 GiB0.01 GiB21±30%
gemma-4-E2B-it-uncensoredQ3_K_S5.1B2.90 GiB0.02 GiB3.70 GiB0.02 GiB21±30%
gemma-4-E2B-it-heretic-araQ3_K_S5.1B2.90 GiB0.02 GiB3.70 GiB0.02 GiB21±30%
gemma-4-E2BQ3_K_S5.1B2.90 GiB0.02 GiB3.70 GiB0.02 GiB21±30%
Holo-3.1-4BI1-Q4_K_M5.2B2.86 GiB0.04 GiB3.70 GiB0.02 GiB21±30%
AfriqueQwen3.5-4BI1-Q4_K_M5.2B2.86 GiB0.04 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 A350M 4GB run?
850 of 2118 indexed open-weight models fit a Arc A350M 4GB at 4,096 context with q4_0 KV cache, the largest being gemma-4-E2B-it-qat-q4_0-unquantized-heretic at I1-IQ3_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a Arc A350M 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 A350M 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.