Qwen · Qwen2.5-Coder · 7.6B parameters

Qwen2.5-Coder 7B VRAM requirements

Qwen2.5-Coder 7B has 28 layers and uses grouped-query attention (4 KV heads). At Q4_K_M the weights come to 4.4 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

5.6 GB of 21.8 GB · 26%

022 GB

Weights 4.4 GB

KV cache 0.4 GB

Runtime overhead 0.8 GB

Qwen2.5-Coder 7B at Q4_K_M leaves 16.1 GB spare on a RTX 4090. There is room to raise the context length or move up a quantisation level.

Generation128tok/s

Prompt processing4547tok/s

Max context128Ktokens

KV per 1K tokens0GB

Every quantisation of Qwen2.5-Coder 7B 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	14.2 GB	15.4 GB	Runs comfortably	124K	41.6
INT8 / W8A8	8.50	7.3 GB	8.5 GB	Runs comfortably	128K	79.4
Q8_0 (GGUF)	8.50	7.3 GB	8.5 GB	Runs comfortably	128K	79.4
FP8 (E4M3)	8.00	7.1 GB	8.3 GB	Runs comfortably	128K	81.5
Q6_K	6.56	5.8 GB	7.1 GB	Runs comfortably	128K	98.5
AWQ 4-bit	4.25	5.3 GB	6.5 GB	Runs comfortably	128K	108
GPTQ 4-bit	4.25	5.3 GB	6.5 GB	Runs comfortably	128K	108
MXFP4	4.25	5.3 GB	6.5 GB	Runs comfortably	128K	108
Q5_K_M	5.67	5.0 GB	6.3 GB	Runs comfortably	128K	113
Q5_K_S	5.52	4.9 GB	6.1 GB	Runs comfortably	128K	116
Q4_K_M	4.85	4.4 GB	5.6 GB	Runs comfortably	128K	128
Q4_K_S	4.58	4.2 GB	5.4 GB	Runs comfortably	128K	134
Q4_0	4.55	4.2 GB	5.4 GB	Runs comfortably	128K	135
IQ4_XS	4.25	3.9 GB	5.2 GB	Runs comfortably	128K	142
Q3_K_M	3.91	3.7 GB	4.9 GB	Runs comfortably	128K	151
IQ3_M	3.70	3.5 GB	4.8 GB	Runs comfortably	128K	158
IQ3_XXS	3.06	3.0 GB	4.3 GB	Runs comfortably	128K	181
Q2_K	2.63	2.7 GB	3.9 GB	Runs comfortably	128K	200
IQ2_XXS	2.06	2.3 GB	3.5 GB	Runs comfortably	128K	234
IQ1_M	1.75	2.0 GB	3.3 GB	Runs comfortably	128K	257

Qwen2.5-Coder 7B 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	469
A100 80GB	80	2039	Runs comfortably	128K	265
RTX 5090	32	1792	Runs comfortably	128K	247
RTX 5080	16	960	Runs comfortably	128K	135
RTX 4090	24	1008	Runs comfortably	128K	128
RTX 5070 Ti	16	896	Runs comfortably	128K	126
RTX 3090	24	936	Runs comfortably	128K	125
Radeon RX 7900 XTX	24	960	Runs comfortably	128K	116
L40S	48	864	Runs comfortably	128K	110
RTX A6000	48	768	Runs comfortably	128K	103
RTX 3080 10GB	10	760	Runs comfortably	62K	102
Mac Studio M3 Ultra 256GB	256	819	Runs comfortably	128K	101
RTX 5070	12	672	Runs comfortably	97K	94.9
RTX 4080 Super	16	736	Runs comfortably	128K	94.4
RTX 4070 Ti Super	16	672	Runs comfortably	128K	86.3
Mac Studio M4 Max 128GB	128	546	Runs comfortably	128K	75.3
RTX 4070 Super	12	504	Runs comfortably	97K	64.9
RTX 4070	12	504	Runs comfortably	97K	64.9
RTX 5060 Ti 16GB	16	448	Runs comfortably	128K	63.6
Arc B580	12	456	Runs comfortably	97K	49.9
RTX 3060 12GB	12	360	Runs comfortably	97K	48.5
NVIDIA DGX Spark (GB10)	128	273	Runs comfortably	128K	38.9
Mac Mini M4 Pro 48GB	48	273	Runs comfortably	128K	37.9
RTX 4060 Ti 16GB	16	288	Runs comfortably	128K	37.3
Ryzen AI Max+ 395 128GB	128	256	Runs comfortably	128K	29.1

Architecture

Parameters	7.6B
Layers	28
Hidden size	3584
Attention heads / KV heads	28 / 4
Head dimension	128
Vocabulary	152,064
Trained context	128K
KV cache per 1K tokens	0 GB
Hugging Face	Qwen/Qwen2.5-Coder-7B-Instruct

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

Qwen2.5-Coder 32B
32.8B · 64 layers Qwen3 0.6B
600M · 28 layers Qwen3 1.7B
1.7B · 28 layers Qwen3 4B
4.0B · 36 layers Qwen3 8B
8.2B · 36 layers Qwen3 14B
14.8B · 40 layers Qwen3 32B
32.8B · 64 layers Qwen3 30B-A3B
30.5B · 3.3B active · 48 layers

Direct answers

Qwen2.5-Coder 7B on RTX 5090
32 GB Qwen2.5-Coder 7B on RTX 4090
24 GB Qwen2.5-Coder 7B on RTX 3090
24 GB Qwen2.5-Coder 7B on RTX 5080
16 GB Qwen2.5-Coder 7B on RTX 5070 Ti
16 GB Qwen2.5-Coder 7B on RTX 5070
12 GB Qwen2.5-Coder 7B on RTX 4070 Ti Super
16 GB Qwen2.5-Coder 7B on RTX 4070 Super
12 GB Qwen2.5-Coder 7B on RTX 4070
12 GB