Qwen · 4.0B parameters

Qwen3 4B VRAM requirements

Qwen3 4B has 36 layers and uses grouped-query attention (8 KV heads). At Q4_K_M the weights come to 2.3 GB, and the best quantisation that fits a 24 GB card is FP16 / BF16.

Model

Hardware

Quantisation

Context 8K

KV cache precision

Runtime

Batch

GPUs

Monitor plugged into this GPU

Runs comfortably

4.2 GB of 21.8 GB · 19%

022 GB

Weights 2.3 GB

KV cache 1.1 GB

Runtime overhead 0.8 GB

Qwen3 4B at Q4_K_M leaves 17.5 GB spare on a RTX 4090. There is room to raise the context length or move up a quantisation level.

Generation202tok/s

Prompt processing8619tok/s

Max context128Ktokens

KV per 1K tokens0GB

Every quantisation of Qwen3 4B 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	7.5 GB	9.4 GB	Runs comfortably	96K	73.9
INT8 / W8A8	8.50	3.9 GB	5.8 GB	Runs comfortably	121K	132
Q8_0 (GGUF)	8.50	3.9 GB	5.8 GB	Runs comfortably	121K	132
FP8 (E4M3)	8.00	3.7 GB	5.7 GB	Runs comfortably	122K	136
Q6_K	6.56	3.1 GB	5.0 GB	Runs comfortably	127K	161
Q5_K_M	5.67	2.7 GB	4.6 GB	Runs comfortably	128K	181
Q5_K_S	5.52	2.6 GB	4.5 GB	Runs comfortably	128K	184
AWQ 4-bit	4.25	2.5 GB	4.4 GB	Runs comfortably	128K	188
GPTQ 4-bit	4.25	2.5 GB	4.4 GB	Runs comfortably	128K	188
MXFP4	4.25	2.5 GB	4.4 GB	Runs comfortably	128K	188
Q4_K_M	4.85	2.3 GB	4.2 GB	Runs comfortably	128K	202
Q4_K_S	4.58	2.2 GB	4.1 GB	Runs comfortably	128K	210
Q4_0	4.55	2.2 GB	4.1 GB	Runs comfortably	128K	211
IQ4_XS	4.25	2.0 GB	4.0 GB	Runs comfortably	128K	221
Q3_K_M	3.91	1.9 GB	3.8 GB	Runs comfortably	128K	233
IQ3_M	3.70	1.8 GB	3.7 GB	Runs comfortably	128K	241
IQ3_XXS	3.06	1.5 GB	3.5 GB	Runs comfortably	128K	270
Q2_K	2.63	1.4 GB	3.3 GB	Runs comfortably	128K	294
IQ2_XXS	2.06	1.1 GB	3.0 GB	Runs comfortably	128K	333
IQ1_M	1.75	1.0 GB	2.9 GB	Runs comfortably	128K	359

Qwen3 4B 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	697
A100 80GB	80	2039	Runs comfortably	128K	407
RTX 5090	32	1792	Runs comfortably	128K	380
RTX 5080	16	960	Runs comfortably	79K	212
RTX 4090	24	1008	Runs comfortably	128K	202
RTX 5070 Ti	16	896	Runs comfortably	79K	198
RTX 3090	24	936	Runs comfortably	128K	196
Radeon RX 7900 XTX	24	960	Runs comfortably	128K	183
L40S	48	864	Runs comfortably	128K	174
RTX A6000	48	768	Runs comfortably	128K	162
RTX 3080 10GB	10	760	Runs comfortably	39K	160
Mac Studio M3 Ultra 256GB	256	819	Runs comfortably	128K	160
RTX 5070	12	672	Runs comfortably	53K	150
RTX 4080 Super	16	736	Runs comfortably	79K	149
RTX 4070 Ti Super	16	672	Runs comfortably	79K	137
Mac Studio M4 Max 128GB	128	546	Runs comfortably	128K	120
RTX 4070 Super	12	504	Runs comfortably	53K	103
RTX 4070	12	504	Runs comfortably	53K	103
RTX 5060 Ti 16GB	16	448	Runs comfortably	79K	101
Arc B580	12	456	Runs comfortably	53K	79.5
RTX 3060 12GB	12	360	Runs comfortably	53K	77.4
NVIDIA DGX Spark (GB10)	128	273	Runs comfortably	128K	62.2
Mac Mini M4 Pro 48GB	48	273	Runs comfortably	128K	60.6
RTX 4060 Ti 16GB	16	288	Runs comfortably	79K	59.6
Ryzen AI Max+ 395 128GB	128	256	Runs comfortably	128K	46.5

Architecture

Parameters	4.0B
Layers	36
Hidden size	2560
Attention heads / KV heads	32 / 8
Head dimension	128
Vocabulary	151,936
Trained context	128K
KV cache per 1K tokens	0 GB
Hugging Face	Qwen/Qwen3-4B

Direct answers

Qwen3 4B on RTX 5090
See the verdict Qwen3 4B on RTX 4090
See the verdict Qwen3 4B on RTX 3090
See the verdict Qwen3 4B on RTX 5080
See the verdict Qwen3 4B on RTX 5070 Ti
See the verdict Qwen3 4B on RTX 5070
See the verdict Qwen3 4B on RTX 4070 Ti Super
See the verdict Qwen3 4B on RTX 4070 Super
See the verdict Qwen3 4B on RTX 4070
See the verdict