gemma-4-31B-it

Google · 30.7B · ctx 256K · apache-2.0 · HuggingFace

gemma-4-31B-it (Google) is a 30.7B parameters open-weight model tracked on AI Hashrate. At 8K context we estimate roughly 24.03 GB VRAM for Q4 (weights 16.89 GB + KV 6.14 GB + 1 GB overhead) and 68.54 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 ≈ 165.8 tok/s (estimated); MI300X 192GB ≈ 109.9 tok/s (estimated); H200 141GB SXM5 ≈ 99.5 tok/s (estimated); H100 80GB SXM5 ≈ 69.4 tok/s (estimated); Gaudi 3 128GB ≈ 67.9 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	165.8 est.	Yes	0.004
MI300X 192GB	192	Q4	109.9 est.	Yes	0.007
H200 141GB SXM5	141	Q4	99.5 est.	Yes	0.003
H100 80GB SXM5	80	Q4	69.4 est.	Yes	0.002
Gaudi 3 128GB	128	Q4	67.9 est.	Yes	0.005
RTX 4090 24GB	24	Q4	59.5 est.	No	0.037
RTX 4090D 24GB (China)	24	Q4	59.5 est.	No	0.037
RTX 3090 Ti 24GB	24	Q4	59.5 est.	No	0.03
RX 7900 XTX 24GB	24	Q4	56.7 est.	No	0.057
RTX 3090 24GB	24	Q4	55.3 est.	No	0.037
B200 192GB SXM	192	FP16	45.6 est.	Yes	0.001
A100 80GB SXM4	80	Q4	42.3 est.	Yes	0.003
RTX 5090 32GB	32	Q4	37.1 est.	Yes	0.019
RX 7900 XT 20GB	20	Q4	32.8 est.	No	0.036
MI300X 192GB	192	FP16	30.2 est.	Yes	0.002
H200 141GB SXM5	141	FP16	27.4 est.	Yes	0.001
RTX 5080 16GB	16	Q4	25.2 est.	No	0.025
RTX 5070 Ti 16GB	16	Q4	23.5 est.	No	0.031
RTX 6000 Ada 48GB	48	Q4	19.9 est.	Yes	0.003
RTX 4080 Super 16GB	16	Q4	19.3 est.	No	0.019
H100 80GB SXM5	80	FP16	19.1 est.	Yes	0.001
Gaudi 3 128GB	128	FP16	18.7 est.	Yes	0.001
L40S 48GB	48	Q4	17.9 est.	Yes	0.002
RTX 4070 Ti Super 16GB	16	Q4	17.6 est.	No	0.022
Mac Studio M3 Ultra 96GB	96	Q4	17.0 est.	Yes	0.004
RX 9070 16GB	16	Q4	16.8 est.	No	0.031
RX 9070 XT 16GB	16	Q4	16.8 est.	No	0.028
Mac Studio M2 Ultra 192GB	192	Q4	16.6 est.	Yes	0.002
RX 7800 XT 16GB	16	Q4	16.4 est.	No	0.033
RTX A6000 48GB	48	Q4	15.9 est.	Yes	0.004
RX 7900 GRE 16GB	16	Q4	15.1 est.	No	0.028
Arc A770 16GB	16	Q4	14.7 est.	No	0.042
RTX 5060 Ti 16GB	16	Q4	11.8 est.	No	0.028
A100 80GB SXM4	80	FP16	11.6 est.	Yes	0.001
MacBook Pro M4 Max 128GB	128	Q4	11.3 est.	Yes	0.003
RTX 5070 12GB	12	Q4	9.9 est.	No	0.018
T4 16GB	16	Q4	8.4 est.	No	0.004
MacBook Pro M3 Max 128GB	128	Q4	8.3 est.	Yes	0.002
RTX 4070 Ti 12GB	12	Q4	8.2 est.	No	0.01
RTX 3080 10GB	10	Q4	7.8 est.	No	0.011
RTX 6000 Ada 48GB	48	FP16	7.7 est.	No	0.001
RTX 4060 Ti 16GB	16	Q4	7.6 est.	No	0.015
RX 7600 XT 16GB	16	Q4	7.6 est.	No	0.023
RTX 4070 12GB	12	Q4	7.4 est.	No	0.012
RTX 4070 Super 12GB	12	Q4	7.4 est.	No	0.012
L40S 48GB	48	FP16	6.9 est.	No	0.001
Arc B580 12GB	12	Q4	6.7 est.	No	0.027
RTX 5090 32GB	32	FP16	6.4 est.	No	0.003
RX 7700 XT 12GB	12	Q4	6.4 est.	No	0.014
RTX A6000 48GB	48	FP16	6.1 est.	No	0.001
Mac mini M4 Pro 64GB	64	Q4	5.7 est.	Yes	0.003
RTX 3060 12GB	12	Q4	5.3 est.	No	0.016
Mac Studio M3 Ultra 96GB	96	FP16	4.7 est.	Yes	0.001
Mac Studio M2 Ultra 192GB	192	FP16	4.6 est.	Yes	0.001
Arc B570 10GB	10	Q4	3.9 est.	No	0.018
Mac mini M4 Pro 64GB	64	FP16	3.9 est.	No	0.002
MacBook Pro M4 Max 128GB	128	FP16	3.1 est.	Yes	0.001
RTX 5060 8GB	8	Q4	2.9 est.	No	0.01
RTX 5060 Ti 8GB	8	Q4	2.9 est.	No	0.008
RTX 3070 8GB	8	Q4	2.9 est.	No	0.006
Mac mini M4 32GB	32	Q4	2.5 est.	Yes	0.003
MacBook Pro M3 Max 128GB	128	FP16	2.3 est.	Yes	0.001
RTX 4090 24GB	24	FP16	2.0 est.	No	0.001
RTX 4090D 24GB (China)	24	FP16	2.0 est.	No	0.001
RTX 3090 Ti 24GB	24	FP16	2.0 est.	No	0.001
RTX 3090 24GB	24	FP16	1.9 est.	No	0.001
RX 7900 XTX 24GB	24	FP16	1.9 est.	No	0.002
RTX 4060 Ti 8GB	8	Q4	1.9 est.	No	0.005
RX 7600 8GB	8	Q4	1.9 est.	No	0.007
RTX 4060 8GB	8	Q4	1.8 est.	No	0.006
RX 7900 XT 20GB	20	FP16	1.1 est.	No	0.001
Mac mini M4 32GB	32	FP16	0.4 est.	No	0.0

Fit checks

• RTX 5090 32GB — fits · ≈37.1 tok/s Q4 est.
• RTX 3090 Ti 24GB — does not fit · ≈59.5 tok/s Q4 est.
• RTX 4090 24GB — does not fit · ≈59.5 tok/s Q4 est.
• RTX 5080 16GB — does not fit · ≈25.2 tok/s Q4 est.
• RTX 6000 Ada 48GB — fits · ≈19.9 tok/s Q4 est.
• RX 7900 XTX 24GB — does not fit · ≈56.7 tok/s Q4 est.
• RTX 3090 24GB — does not fit · ≈55.3 tok/s Q4 est.
• RTX 5070 Ti 16GB — does not fit · ≈23.5 tok/s Q4 est.
• L40S 48GB — fits · ≈17.9 tok/s Q4 est.
• Mac Studio M3 Ultra 96GB — fits · ≈17.0 tok/s Q4 est.
• Mac Studio M2 Ultra 192GB — fits · ≈16.6 tok/s Q4 est.
• RX 7900 XT 20GB — does not fit · ≈32.8 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-4-26B-A4B-it

All Gemma models →