Qwen · 30.5B parameters · 3.3B active

Qwen3 30B-A3B VRAM requirements

Qwen3 30B-A3B has 48 layers and uses grouped-query attention (4 KV heads). At Q4_K_M the weights come to 17.3 GB, and the best quantisation that fits a 24 GB card is Q5_K_M.

Model

Hardware

Quantisation

Context 8K

KV cache precision

Runtime

Batch

GPUs

Monitor plugged into this GPU

Runs comfortably

18.8 GB of 21.8 GB · 86%

022 GB

Weights 17.3 GB

KV cache 0.8 GB

Runtime overhead 0.8 GB

Qwen3 30B-A3B at Q4_K_M leaves 2.9 GB spare on a RTX 4090. There is room to raise the context length or move up a quantisation level.

Generation118tok/s

Prompt processing10374tok/s

Max context39Ktokens

KV per 1K tokens0GB

Every quantisation of Qwen3 30B-A3B on a RTX 4090

Highlighted row is the highest quality that still fits at 8K context.

Quantisation	bpw	Weights	Total	Verdict	Max ctx	tok/s
FP16 / BF16	16.00	56.9 GB	58.4 GB	Won't fit	—	8.34
INT8 / W8A8	8.50	30.1 GB	31.6 GB	Won't fit	—	24.6
Q8_0 (GGUF)	8.50	30.1 GB	31.6 GB	Won't fit	—	24.6
FP8 (E4M3)	8.00	28.4 GB	30.0 GB	Won't fit	—	28.3
Q6_K	6.56	23.3 GB	24.9 GB	Won't fit	—	49.9
Q5_K_M	5.67	20.2 GB	21.7 GB	Fits, but tight	9K	111
Q5_K_S	5.52	19.6 GB	21.2 GB	Fits, but tight	14K	112
Q4_K_M	4.85	17.3 GB	18.8 GB	Runs comfortably	39K	118
Q4_K_S	4.58	16.3 GB	17.9 GB	Runs comfortably	49K	121
Q4_0	4.55	16.2 GB	17.8 GB	Runs comfortably	50K	122
AWQ 4-bit	4.25	16.0 GB	17.5 GB	Runs comfortably	54K	123
GPTQ 4-bit	4.25	16.0 GB	17.5 GB	Runs comfortably	54K	123
MXFP4	4.25	16.0 GB	17.5 GB	Runs comfortably	54K	123
IQ4_XS	4.25	15.2 GB	16.7 GB	Runs comfortably	62K	125
Q3_K_M	3.91	14.0 GB	15.5 GB	Runs comfortably	74K	129
IQ3_M	3.70	13.3 GB	14.8 GB	Runs comfortably	82K	132
IQ3_XXS	3.06	11.1 GB	12.6 GB	Runs comfortably	106K	140
Q2_K	2.63	9.6 GB	11.1 GB	Runs comfortably	122K	147
IQ2_XXS	2.06	7.6 GB	9.1 GB	Runs comfortably	128K	156
IQ1_M	1.75	6.5 GB	8.0 GB	Runs comfortably	128K	162

Qwen3 30B-A3B on each GPU

Q4_K_M weights at 8K context, single card, monitor attached.

GPU	VRAM	GB/s	Verdict	Max ctx	tok/s
H100 SXM 80GB	80	3350	Runs comfortably	128K	192
A100 80GB	80	2039	Runs comfortably	128K	163
RTX 5090	32	1792	Runs comfortably	120K	159
RTX 4090	24	1008	Runs comfortably	39K	118
RTX 3090	24	936	Runs comfortably	39K	117
Radeon RX 7900 XTX	24	960	Runs comfortably	39K	112
L40S	48	864	Runs comfortably	128K	109
RTX A6000	48	768	Runs comfortably	128K	105
Mac Studio M3 Ultra 256GB	256	819	Runs comfortably	128K	104
Mac Studio M4 Max 128GB	128	546	Runs comfortably	128K	86.4
NVIDIA DGX Spark (GB10)	128	273	Runs comfortably	128K	53.5
Mac Mini M4 Pro 48GB	48	273	Runs comfortably	128K	52.4
RTX 5080	16	960	Won't fit	—	46.2
RTX 5070 Ti	16	896	Won't fit	—	45.7
Ryzen AI Max+ 395 128GB	128	256	Runs comfortably	128K	42.2
RTX 4080 Super	16	736	Won't fit	—	40.9
RTX 4070 Ti Super	16	672	Won't fit	—	40.2
RTX 5060 Ti 16GB	16	448	Won't fit	—	39.8
RTX 4060 Ti 16GB	16	288	Won't fit	—	32.0
RTX 5070	12	672	Won't fit	—	29.5
RTX 4070 Super	12	504	Won't fit	—	26.3
RTX 4070	12	504	Won't fit	—	26.3
RTX 3060 12GB	12	360	Won't fit	—	26.1
RTX 3080 10GB	10	760	Won't fit	—	24.7
Arc B580	12	456	Won't fit	—	22.4

Architecture

Parameters	30.5B
Active per token	3.3B of 128 experts, top-8
Layers	48
Hidden size	2048
Attention heads / KV heads	32 / 4
Head dimension	128
Vocabulary	151,936
Trained context	128K
KV cache per 1K tokens	0 GB
Hugging Face	Qwen/Qwen3-30B-A3B

Direct answers

Qwen3 30B-A3B on RTX 5090
See the verdict Qwen3 30B-A3B on RTX 4090
See the verdict Qwen3 30B-A3B on RTX 3090
See the verdict Qwen3 30B-A3B on RTX 5080
See the verdict Qwen3 30B-A3B on RTX 5070 Ti
See the verdict Qwen3 30B-A3B on RTX 5070
See the verdict Qwen3 30B-A3B on RTX 4070 Ti Super
See the verdict Qwen3 30B-A3B on RTX 4070 Super
See the verdict Qwen3 30B-A3B on RTX 4070
See the verdict