Qwen2.5-32B-Instruct

Alibaba · 32.8B · ctx 128K · apache-2.0 · HuggingFace

Qwen2.5-32B-Instruct (Alibaba) is a 32.8B parameters open-weight model tracked on AI Hashrate. At 8K context we estimate roughly 25.6 GB VRAM for Q4 (weights 18.04 GB + KV 6.56 GB + 1 GB overhead) and 73.16 GB for FP16. About 16 GPUs/accelerators in our catalog fully fit this model at Q4 under that context assumption. Fastest Q4 configs: B200 192GB SXM ≈ 155.2 tok/s (estimated); MI300X 192GB ≈ 102.8 tok/s (estimated); H200 141GB SXM5 ≈ 93.1 tok/s (estimated); RTX 5090 32GB ≈ 75.0 tok/s (measured); H100 80GB SXM5 ≈ 65.0 tok/s (estimated). Use the table below for VRAM fit and tok/s per dollar (list MSRP). Estimates follow memory-bandwidth math; measured rows override when present. See methodology for details. Methodology.

72 hardware configs — sorted by tok/s. Default fit context 8K.

Hardware	VRAM	Quant	tok/s	Fits?	tok/s/$
B200 192GB SXM	192	Q4	155.2 est.	Yes	0.004
MI300X 192GB	192	Q4	102.8 est.	Yes	0.007
H200 141GB SXM5	141	Q4	93.1 est.	Yes	0.003
RTX 5090 32GB	32	Q4	75.0 measured	Yes	0.038
H100 80GB SXM5	80	Q4	65.0 est.	Yes	0.002
Gaudi 3 128GB	128	Q4	63.6 est.	Yes	0.005
A100 80GB SXM4	80	Q4	55.83 measured	Yes	0.004
RTX 4090D 24GB (China)	24	Q4	49.1 est.	No	0.031
RTX 3090 Ti 24GB	24	Q4	49.1 est.	No	0.025
B200 192GB SXM	192	FP16	42.7 est.	Yes	0.001
RTX 4090 24GB	24	Q4	42.0 measured	No	0.026
RX 7900 XTX 24GB	24	Q4	30.0 measured	No	0.03
MI300X 192GB	192	FP16	28.3 est.	Yes	0.002
RTX 3090 24GB	24	Q4	28.0 measured	No	0.019
RX 7900 XT 20GB	20	Q4	27.1 est.	No	0.03
H200 141GB SXM5	141	FP16	25.6 est.	Yes	0.001
RTX 5080 16GB	16	Q4	20.8 est.	No	0.021
RTX 5070 Ti 16GB	16	Q4	19.4 est.	No	0.026
RTX 6000 Ada 48GB	48	Q4	18.6 est.	Yes	0.003
H100 80GB SXM5	80	FP16	17.9 est.	Yes	0.001
Gaudi 3 128GB	128	FP16	17.5 est.	Yes	0.001
L40S 48GB	48	Q4	16.8 est.	Yes	0.002
RTX 4080 Super 16GB	16	Q4	15.9 est.	No	0.016
Mac Studio M3 Ultra 96GB	96	Q4	15.9 est.	Yes	0.004
Mac Studio M2 Ultra 192GB	192	Q4	15.5 est.	Yes	0.002
RTX A6000 48GB	48	Q4	14.9 est.	Yes	0.003
RTX 4070 Ti Super 16GB	16	Q4	14.6 est.	No	0.018
RX 9070 16GB	16	Q4	13.9 est.	No	0.025
RX 9070 XT 16GB	16	Q4	13.9 est.	No	0.023
RX 7800 XT 16GB	16	Q4	13.5 est.	No	0.027
RX 7900 GRE 16GB	16	Q4	12.5 est.	No	0.023
Arc A770 16GB	16	Q4	12.1 est.	No	0.035
A100 80GB SXM4	80	FP16	10.9 est.	Yes	0.001
MacBook Pro M4 Max 128GB	128	Q4	10.6 est.	Yes	0.002
RTX 5060 Ti 16GB	16	Q4	9.7 est.	No	0.023
RTX 5070 12GB	12	Q4	8.2 est.	No	0.015
MacBook Pro M3 Max 128GB	128	Q4	7.8 est.	Yes	0.002
T4 16GB	16	Q4	6.9 est.	No	0.003
RTX 4070 Ti 12GB	12	Q4	6.7 est.	No	0.008
RTX 3080 10GB	10	Q4	6.4 est.	No	0.009
RTX 6000 Ada 48GB	48	FP16	6.3 est.	No	0.001
RTX 4060 Ti 16GB	16	Q4	6.2 est.	No	0.012
RX 7600 XT 16GB	16	Q4	6.2 est.	No	0.019
RTX 4070 12GB	12	Q4	6.1 est.	No	0.01
RTX 4070 Super 12GB	12	Q4	6.1 est.	No	0.01
L40S 48GB	48	FP16	5.7 est.	No	0.001
Arc B580 12GB	12	Q4	5.6 est.	No	0.022
RX 7700 XT 12GB	12	Q4	5.3 est.	No	0.012
Mac mini M4 Pro 64GB	64	Q4	5.3 est.	Yes	0.002
RTX 5090 32GB	32	FP16	5.2 est.	No	0.003
RTX A6000 48GB	48	FP16	5.0 est.	No	0.001
RTX 3060 12GB	12	Q4	4.4 est.	No	0.013
Mac Studio M3 Ultra 96GB	96	FP16	4.4 est.	Yes	0.001
Mac Studio M2 Ultra 192GB	192	FP16	4.3 est.	Yes	0.001
Arc B570 10GB	10	Q4	3.2 est.	No	0.015
Mac mini M4 Pro 64GB	64	FP16	3.2 est.	No	0.001
MacBook Pro M4 Max 128GB	128	FP16	2.9 est.	Yes	0.001
RTX 5060 8GB	8	Q4	2.4 est.	No	0.008
RTX 5060 Ti 8GB	8	Q4	2.4 est.	No	0.006
RTX 3070 8GB	8	Q4	2.4 est.	No	0.005
Mac mini M4 32GB	32	Q4	2.3 est.	Yes	0.002
MacBook Pro M3 Max 128GB	128	FP16	2.1 est.	Yes	0.001
RTX 4090 24GB	24	FP16	1.7 est.	No	0.001
RTX 4090D 24GB (China)	24	FP16	1.7 est.	No	0.001
RTX 3090 Ti 24GB	24	FP16	1.7 est.	No	0.001
RX 7900 XTX 24GB	24	FP16	1.6 est.	No	0.002
RTX 4060 Ti 8GB	8	Q4	1.6 est.	No	0.004
RX 7600 8GB	8	Q4	1.6 est.	No	0.006
RTX 3090 24GB	24	FP16	1.5 est.	No	0.001
RTX 4060 8GB	8	Q4	1.5 est.	No	0.005
RX 7900 XT 20GB	20	FP16	0.9 est.	No	0.001
Mac mini M4 32GB	32	FP16	0.3 est.	No	0.0

Fit checks

• RTX 5090 32GB — fits · ≈75.0 tok/s Q4 measured
• RTX 3090 Ti 24GB — does not fit · ≈49.1 tok/s Q4 est.
• RTX 4090 24GB — does not fit · ≈42.0 tok/s Q4 measured
• RTX 5080 16GB — does not fit · ≈20.8 tok/s Q4 est.
• RTX 6000 Ada 48GB — fits · ≈18.6 tok/s Q4 est.
• RX 7900 XTX 24GB — does not fit · ≈30.0 tok/s Q4 measured
• RTX 3090 24GB — does not fit · ≈28.0 tok/s Q4 measured
• RTX 5070 Ti 16GB — does not fit · ≈19.4 tok/s Q4 est.
• L40S 48GB — fits · ≈16.8 tok/s Q4 est.
• Mac Studio M3 Ultra 96GB — fits · ≈15.9 tok/s Q4 est.
• Mac Studio M2 Ultra 192GB — fits · ≈15.5 tok/s Q4 est.
• RX 7900 XT 20GB — does not fit · ≈27.1 tok/s Q4 est.

12 popular retail GPUs; every row in the table above also links its Fits? verdict to the full fit check.

Same family: Qwen

Qwen3 Qwen3-8B · Qwen3-14B
Qwen3 2507 Qwen3-4B-Instruct-2507 · Qwen3-4B-Thinking-2507 · Qwen3-30B-A3B-Instruct-2507 · Qwen3-235B-A22B-Instruct-2507
Qwen2.5 Qwen2.5-7B-Instruct · Qwen2.5-14B-Instruct · Qwen2.5-72B-Instruct
Qwen2.5 Coder Qwen2.5-Coder-32B-Instruct
Qwen3 Next Qwen3-Next-80B-A3B-Instruct
Qwen3 Coder Qwen3-Coder-480B-A35B-Instruct · Qwen3-Coder-Next
Qwen3.5 Qwen3.5-2B · Qwen3.5-9B · Qwen3.5-27B · Qwen3.5-35B-A3B · Qwen3.5-122B-A10B · Qwen3.5-397B-A17B
Qwen3.6 Qwen3.6-27B · Qwen3.6-35B-A3B

All Qwen models →