Gemma · Gemma 3 · 4.3B parameters

Gemma 3 4B VRAM requirements

Gemma 3 4B has 34 layers and uses grouped-query attention (4 KV heads). At Q4_K_M the weights come to 2.5 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

3.6 GB of 21.8 GB · 16%

022 GB

Weights 2.5 GB

KV cache 0.3 GB

Runtime overhead 0.8 GB

Gemma 3 4B at Q4_K_M leaves 18.2 GB spare on a RTX 4090. There is room to raise the context length or move up a quantisation level.

Generation215tok/s

Prompt processing8058tok/s

Max context128Ktokens

KV per 1K tokens0GB

Every quantisation of Gemma 3 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	8.0 GB	9.1 GB	Runs comfortably	128K	72.5
INT8 / W8A8	8.50	4.1 GB	5.2 GB	Runs comfortably	128K	136
Q8_0 (GGUF)	8.50	4.1 GB	5.2 GB	Runs comfortably	128K	136
FP8 (E4M3)	8.00	4.0 GB	5.1 GB	Runs comfortably	128K	140
Q6_K	6.56	3.3 GB	4.4 GB	Runs comfortably	128K	167
AWQ 4-bit	4.25	3.0 GB	4.1 GB	Runs comfortably	128K	179
GPTQ 4-bit	4.25	3.0 GB	4.1 GB	Runs comfortably	128K	179
MXFP4	4.25	3.0 GB	4.1 GB	Runs comfortably	128K	179
Q5_K_M	5.67	2.8 GB	3.9 GB	Runs comfortably	128K	191
Q5_K_S	5.52	2.8 GB	3.9 GB	Runs comfortably	128K	196
Q4_K_M	4.85	2.5 GB	3.6 GB	Runs comfortably	128K	215
Q4_K_S	4.58	2.4 GB	3.5 GB	Runs comfortably	128K	225
Q4_0	4.55	2.4 GB	3.4 GB	Runs comfortably	128K	226
IQ4_XS	4.25	2.2 GB	3.3 GB	Runs comfortably	128K	237
Q3_K_M	3.91	2.1 GB	3.2 GB	Runs comfortably	128K	251
IQ3_M	3.70	2.0 GB	3.1 GB	Runs comfortably	128K	261
IQ3_XXS	3.06	1.7 GB	2.8 GB	Runs comfortably	128K	295
Q2_K	2.63	1.5 GB	2.6 GB	Runs comfortably	128K	324
IQ2_XXS	2.06	1.3 GB	2.4 GB	Runs comfortably	128K	372
IQ1_M	1.75	1.2 GB	2.3 GB	Runs comfortably	128K	404

Gemma 3 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	742
A100 80GB	80	2039	Runs comfortably	128K	434
RTX 5090	32	1792	Runs comfortably	128K	405
RTX 5080	16	960	Runs comfortably	128K	226
RTX 4090	24	1008	Runs comfortably	128K	215
RTX 5070 Ti	16	896	Runs comfortably	128K	211
RTX 3090	24	936	Runs comfortably	128K	209
Radeon RX 7900 XTX	24	960	Runs comfortably	128K	196
L40S	48	864	Runs comfortably	128K	186
RTX A6000	48	768	Runs comfortably	128K	173
RTX 3080 10GB	10	760	Runs comfortably	128K	171
Mac Studio M3 Ultra 256GB	256	819	Runs comfortably	128K	170
RTX 5070	12	672	Runs comfortably	128K	160
RTX 4080 Super	16	736	Runs comfortably	128K	159
RTX 4070 Ti Super	16	672	Runs comfortably	128K	146
Mac Studio M4 Max 128GB	128	546	Runs comfortably	128K	128
RTX 4070 Super	12	504	Runs comfortably	128K	110
RTX 4070	12	504	Runs comfortably	128K	110
RTX 5060 Ti 16GB	16	448	Runs comfortably	128K	108
Arc B580	12	456	Runs comfortably	128K	84.8
RTX 3060 12GB	12	360	Runs comfortably	128K	82.6
NVIDIA DGX Spark (GB10)	128	273	Runs comfortably	128K	66.4
Mac Mini M4 Pro 48GB	48	273	Runs comfortably	128K	64.6
RTX 4060 Ti 16GB	16	288	Runs comfortably	128K	63.6
Ryzen AI Max+ 395 128GB	128	256	Runs comfortably	128K	49.6

Architecture

Parameters	4.3B
Layers	34
Hidden size	2560
Attention heads / KV heads	8 / 4
Head dimension	256
Vocabulary	262,144
Trained context	128K
Sliding window	1024 (every 6th layer is global)
KV cache per 1K tokens	0 GB
Hugging Face	google/gemma-3-4b-it

The Gemma family

Google's open models. Two things dominate the memory: a 262k vocabulary — on Gemma 3 1B the embedding table is about a third of the file — and sliding-window attention, where only every sixth layer of Gemma 3 sees the full context, and every second layer on Gemma 2. The cache grows far more slowly than the context length suggests.

huggingface.co/google · ai.google.dev/gemma · all 6 Gemma models

Gemma 3 1B
1B · 26 layers Gemma 3 12B
12.2B · 48 layers Gemma 3 27B
27.4B · 62 layers Gemma 2 9B
9.2B · 42 layers Gemma 2 27B
27.2B · 46 layers

Direct answers

Gemma 3 4B on RTX 5090
See the verdict Gemma 3 4B on RTX 4090
See the verdict Gemma 3 4B on RTX 3090
See the verdict Gemma 3 4B on RTX 5080
See the verdict Gemma 3 4B on RTX 5070 Ti
See the verdict Gemma 3 4B on RTX 5070
See the verdict Gemma 3 4B on RTX 4070 Ti Super
See the verdict Gemma 3 4B on RTX 4070 Super
See the verdict Gemma 3 4B on RTX 4070
See the verdict