Qwen · Qwen3 · 8.2B parameters

Qwen3 8B VRAM requirements

Qwen3 8B has 36 layers and uses grouped-query attention (8 KV heads). At Q4_K_M the weights come to 4.7 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

6.7 GB of 21.8 GB · 31%

022 GB

Weights 4.7 GB

KV cache 1.1 GB

Runtime overhead 0.8 GB

Qwen3 8B at Q4_K_M leaves 15.1 GB spare on a RTX 4090. There is room to raise the context length or move up a quantisation level.

Generation112tok/s

Prompt processing4231tok/s

Max context115Ktokens

KV per 1K tokens0GB

Every quantisation of Qwen3 8B 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	15.3 GB	17.2 GB	Runs comfortably	40K	37.9
INT8 / W8A8	8.50	7.8 GB	9.8 GB	Runs comfortably	93K	71.0
Q8_0 (GGUF)	8.50	7.8 GB	9.8 GB	Runs comfortably	93K	71.0
FP8 (E4M3)	8.00	7.6 GB	9.6 GB	Runs comfortably	95K	72.6
Q6_K	6.56	6.3 GB	8.2 GB	Runs comfortably	104K	87.0
AWQ 4-bit	4.25	5.8 GB	7.7 GB	Runs comfortably	108K	93.8
GPTQ 4-bit	4.25	5.8 GB	7.7 GB	Runs comfortably	108K	93.8
MXFP4	4.25	5.8 GB	7.7 GB	Runs comfortably	108K	93.8
Q5_K_M	5.67	5.4 GB	7.4 GB	Runs comfortably	110K	99.1
Q5_K_S	5.52	5.3 GB	7.2 GB	Runs comfortably	111K	101
Q4_K_M	4.85	4.7 GB	6.7 GB	Runs comfortably	115K	112
Q4_K_S	4.58	4.5 GB	6.4 GB	Runs comfortably	117K	116
Q4_0	4.55	4.5 GB	6.4 GB	Runs comfortably	117K	117
IQ4_XS	4.25	4.2 GB	6.2 GB	Runs comfortably	119K	123
Q3_K_M	3.91	4.0 GB	5.9 GB	Runs comfortably	121K	130
IQ3_M	3.70	3.8 GB	5.7 GB	Runs comfortably	122K	135
IQ3_XXS	3.06	3.3 GB	5.2 GB	Runs comfortably	126K	152
Q2_K	2.63	2.9 GB	4.9 GB	Runs comfortably	128K	167
IQ2_XXS	2.06	2.5 GB	4.4 GB	Runs comfortably	128K	192
IQ1_M	1.75	2.2 GB	4.2 GB	Runs comfortably	128K	208

Qwen3 8B 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	405
A100 80GB	80	2039	Runs comfortably	128K	230
RTX 5090	32	1792	Runs comfortably	128K	214
RTX 5080	16	960	Runs comfortably	62K	117
RTX 4090	24	1008	Runs comfortably	115K	112
RTX 5070 Ti	16	896	Runs comfortably	62K	109
RTX 3090	24	936	Runs comfortably	115K	108
Radeon RX 7900 XTX	24	960	Runs comfortably	115K	101
L40S	48	864	Runs comfortably	128K	96.1
RTX A6000	48	768	Runs comfortably	128K	89.3
RTX 3080 10GB	10	760	Runs comfortably	22K	88.4
Mac Studio M3 Ultra 256GB	256	819	Runs comfortably	128K	87.9
RTX 5070	12	672	Runs comfortably	35K	82.6
RTX 4080 Super	16	736	Runs comfortably	62K	82.1
RTX 4070 Ti Super	16	672	Runs comfortably	62K	75.1
Mac Studio M4 Max 128GB	128	546	Runs comfortably	128K	65.6
RTX 4070 Super	12	504	Runs comfortably	35K	56.5
RTX 4070	12	504	Runs comfortably	35K	56.5
RTX 5060 Ti 16GB	16	448	Runs comfortably	62K	55.4
Arc B580	12	456	Runs comfortably	35K	43.4
RTX 3060 12GB	12	360	Runs comfortably	35K	42.3
NVIDIA DGX Spark (GB10)	128	273	Runs comfortably	128K	33.9
Mac Mini M4 Pro 48GB	48	273	Runs comfortably	128K	33.0
RTX 4060 Ti 16GB	16	288	Runs comfortably	62K	32.5
Ryzen AI Max+ 395 128GB	128	256	Runs comfortably	128K	25.3

Architecture

Parameters	8.2B
Layers	36
Hidden size	4096
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/Qwen3-8B

The Qwen family

Alibaba's series, and the broadest size ladder available — Qwen3 runs from 0.6B to 32B dense, plus 30B-A3B and 235B-A22B as mixture-of-experts. The 152k vocabulary makes the embedding table a large share of a small model's file. Qwen2.5-Coder is the same architecture trained for code.

huggingface.co/Qwen · qwenlm.github.io · all 14 Qwen models

Qwen3 0.6B
600M · 28 layers Qwen3 1.7B
1.7B · 28 layers Qwen3 4B
4.0B · 36 layers Qwen3 14B
14.8B · 40 layers Qwen3 32B
32.8B · 64 layers Qwen3 30B-A3B
30.5B · 3.3B active · 48 layers Qwen3 235B-A22B
235.1B · 22.1B active · 94 layers Qwen2.5 7B
7.6B · 28 layers

Direct answers

Qwen3 8B on RTX 5090
See the verdict Qwen3 8B on RTX 4090
See the verdict Qwen3 8B on RTX 3090
See the verdict Qwen3 8B on RTX 5080
See the verdict Qwen3 8B on RTX 5070 Ti
See the verdict Qwen3 8B on RTX 5070
See the verdict Qwen3 8B on RTX 4070 Ti Super
See the verdict Qwen3 8B on RTX 4070 Super
See the verdict Qwen3 8B on RTX 4070
See the verdict