gemma-4-26B-A4B-it — which GPUs fit

Google · 25.2B (active 3.8B) · MoE · ctx 256K · apache-2.0 · HuggingFace

gemma-4-26B-A4B-it (Google) is a 25.2B parameters MoE (3.8B active per token) open-weight model tracked on AI Hashrate. At 8K context we estimate roughly 15.62 GB VRAM for Q4 (weights 13.86 GB + KV 0.76 GB + 1 GB overhead) and 52.16 GB for FP16. About 22 GPUs/accelerators in our catalog fully fit this model at Q4 under that context assumption. Fastest Q4 configs: B200 192GB SXM ≈ 1339.7 tok/s (estimated); MI300X 192GB ≈ 887.6 tok/s (estimated); H200 141GB SXM5 ≈ 803.8 tok/s (estimated); H100 80GB SXM5 ≈ 561.0 tok/s (estimated); Gaudi 3 128GB ≈ 548.6 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.

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

Hardware	VRAM	Quant	tok/s	Fits?	tok/s/$
B200 192GB SXM	192	Q4	1339.7 est.	Yes	0.033
MI300X 192GB	192	Q4	887.6 est.	Yes	0.059
H200 141GB SXM5	141	Q4	803.8 est.	Yes	0.023
H100 80GB SXM5	80	Q4	561.0 est.	Yes	0.019
Gaudi 3 128GB	128	Q4	548.6 est.	Yes	0.044
RTX 5080 16GB	16	Q4	482.0 est.	No	0.482
RTX 5070 Ti 16GB	16	Q4	449.8 est.	No	0.601
RTX 4080 Super 16GB	16	Q4	369.5 est.	No	0.37
B200 192GB SXM	192	FP16	368.4 est.	Yes	0.009
A100 80GB SXM4	80	Q4	341.5 est.	Yes	0.023
RTX 4070 Ti Super 16GB	16	Q4	337.4 est.	No	0.422
RX 9070 16GB	16	Q4	321.3 est.	No	0.585
RX 9070 XT 16GB	16	Q4	321.3 est.	No	0.536
RX 7800 XT 16GB	16	Q4	313.3 est.	No	0.628
RTX 5090 32GB	32	Q4	300.1 est.	Yes	0.15
RX 7900 GRE 16GB	16	Q4	289.2 est.	No	0.527
Arc A770 16GB	16	Q4	281.1 est.	No	0.805
MI300X 192GB	192	FP16	244.1 est.	Yes	0.016
RTX 5060 Ti 16GB	16	Q4	224.9 est.	No	0.524
H200 141GB SXM5	141	FP16	221.1 est.	Yes	0.006
RTX 5070 12GB	12	Q4	189.8 est.	No	0.346
RTX 4090 24GB	24	Q4	168.8 est.	Yes	0.106
RTX 4090D 24GB (China)	24	Q4	168.8 est.	Yes	0.106
RTX 3090 Ti 24GB	24	Q4	168.8 est.	Yes	0.084
RX 7900 XTX 24GB	24	Q4	160.8 est.	Yes	0.161
RTX 6000 Ada 48GB	48	Q4	160.8 est.	Yes	0.024
T4 16GB	16	Q4	160.7 est.	No	0.08
RTX 3090 24GB	24	Q4	156.7 est.	Yes	0.105
RTX 4070 Ti 12GB	12	Q4	156.4 est.	No	0.196
H100 80GB SXM5	80	FP16	154.3 est.	Yes	0.005
Gaudi 3 128GB	128	FP16	150.9 est.	Yes	0.012
RTX 3080 10GB	10	Q4	149.0 est.	No	0.213
L40S 48GB	48	Q4	144.7 est.	Yes	0.018
RTX 4060 Ti 16GB	16	Q4	144.6 est.	No	0.29
RX 7600 XT 16GB	16	Q4	144.6 est.	No	0.44
RTX 4070 12GB	12	Q4	142.3 est.	No	0.238
RTX 4070 Super 12GB	12	Q4	142.3 est.	No	0.238
Mac Studio M3 Ultra 96GB	96	Q4	137.2 est.	Yes	0.034
Mac Studio M2 Ultra 192GB	192	Q4	134.0 est.	Yes	0.019
RX 7900 XT 20GB	20	Q4	134.0 est.	Yes	0.149
Arc B580 12GB	12	Q4	128.8 est.	No	0.517
RTX A6000 48GB	48	Q4	128.6 est.	Yes	0.029
RX 7700 XT 12GB	12	Q4	122.0 est.	No	0.272
RTX 6000 Ada 48GB	48	FP16	107.0 est.	No	0.016
RTX 3060 12GB	12	Q4	101.7 est.	No	0.309
L40S 48GB	48	FP16	96.3 est.	No	0.012
A100 80GB SXM4	80	FP16	93.9 est.	Yes	0.006
MacBook Pro M4 Max 128GB	128	Q4	91.4 est.	Yes	0.021
RTX 5090 32GB	32	FP16	88.7 est.	No	0.044
RTX A6000 48GB	48	FP16	85.6 est.	No	0.019
Arc B570 10GB	10	Q4	74.5 est.	No	0.34
MacBook Pro M3 Max 128GB	128	Q4	67.0 est.	Yes	0.016
RTX 5060 8GB	8	Q4	56.2 est.	No	0.188
RTX 5060 Ti 8GB	8	Q4	56.2 est.	No	0.148
RTX 3070 8GB	8	Q4	56.2 est.	No	0.113
Mac mini M4 Pro 64GB	64	Q4	45.7 est.	Yes	0.021
Mac Studio M3 Ultra 96GB	96	FP16	37.7 est.	Yes	0.009
Mac Studio M2 Ultra 192GB	192	FP16	36.8 est.	Yes	0.005
RTX 4060 Ti 8GB	8	Q4	36.1 est.	No	0.09
RX 7600 8GB	8	Q4	36.1 est.	No	0.134
RTX 4060 8GB	8	Q4	34.1 est.	No	0.114
RTX 4090 24GB	24	FP16	28.1 est.	No	0.018
RTX 4090D 24GB (China)	24	FP16	28.1 est.	No	0.018
RTX 3090 Ti 24GB	24	FP16	28.1 est.	No	0.014
RX 7900 XTX 24GB	24	FP16	26.7 est.	No	0.027
RTX 3090 24GB	24	FP16	26.1 est.	No	0.017
MacBook Pro M4 Max 128GB	128	FP16	25.1 est.	Yes	0.006
Mac mini M4 32GB	32	Q4	20.1 est.	Yes	0.02
MacBook Pro M3 Max 128GB	128	FP16	18.4 est.	Yes	0.004
RX 7900 XT 20GB	20	FP16	15.5 est.	No	0.017
Mac mini M4 Pro 64GB	64	FP16	12.6 est.	Yes	0.006
RTX 5080 16GB	16	FP16	11.9 est.	No	0.012
RTX 5070 Ti 16GB	16	FP16	11.1 est.	No	0.015
RTX 4080 Super 16GB	16	FP16	9.1 est.	No	0.009
RTX 4070 Ti Super 16GB	16	FP16	8.3 est.	No	0.01
RX 9070 16GB	16	FP16	7.9 est.	No	0.014
RX 9070 XT 16GB	16	FP16	7.9 est.	No	0.013
RX 7800 XT 16GB	16	FP16	7.7 est.	No	0.015
RX 7900 GRE 16GB	16	FP16	7.1 est.	No	0.013
Arc A770 16GB	16	FP16	6.9 est.	No	0.02
Mac mini M4 32GB	32	FP16	5.9 est.	No	0.006
RTX 5060 Ti 16GB	16	FP16	5.5 est.	No	0.013
T4 16GB	16	FP16	4.0 est.	No	0.002
RTX 4060 Ti 16GB	16	FP16	3.6 est.	No	0.007
RX 7600 XT 16GB	16	FP16	3.6 est.	No	0.011

Fit checks

RTX 5090 32GB — fits · ≈300.1 tok/s Q4 est.
RTX 3090 Ti 24GB — fits · ≈168.8 tok/s Q4 est.
RTX 4090 24GB — fits · ≈168.8 tok/s Q4 est.
RTX 5080 16GB — does not fit · ≈482.0 tok/s Q4 est.
RTX 6000 Ada 48GB — fits · ≈160.8 tok/s Q4 est.
RX 7900 XTX 24GB — fits · ≈160.8 tok/s Q4 est.
RTX 3090 24GB — fits · ≈156.7 tok/s Q4 est.
RTX 5070 Ti 16GB — does not fit · ≈449.8 tok/s Q4 est.
L40S 48GB — fits · ≈144.7 tok/s Q4 est.
Mac Studio M3 Ultra 96GB — fits · ≈137.2 tok/s Q4 est.
Mac Studio M2 Ultra 192GB — fits · ≈134.0 tok/s Q4 est.
RX 7900 XT 20GB — fits · ≈134.0 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: Gemma

Gemma 4 gemma-4-E2B-it · gemma-4-E4B-it · gemma-4-12b-it · gemma-4-31B-it

All Gemma models →