Skip to content

Gemma 4 12B VRAM requirements

How much VRAM Gemma 4 12B needs
4-bit weights: 6.8 GBWith 8K context: 8.2 GBWith 32K context: 8.4 GB8-bit, 8K context: 13.3 GB
Parameters
12B
Longest context
256K
Released
May 2026
License
Apache 2.0
Strongest at
Chat, Coding, Docs

How much VRAM Gemma 4 12B needs

Memory Gemma 4 12B needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit22.3 GB23.7 GB23.9 GB24.7 GB
8-bit11.8 GB13.3 GB13.5 GB14.2 GB
6-bit9.1 GB10.6 GB10.8 GB11.5 GB
5-bit7.9 GB9.4 GB9.6 GB10.3 GB
4-bit6.8 GB8.2 GB8.4 GB9.1 GB
3-bit5.4 GB6.9 GB7.1 GB7.8 GB
2-bit4.2 GB5.6 GB5.8 GB6.6 GB
1.6-bit2.3 GB3.7 GB3.9 GB4.7 GB

One conversation, 16-bit context cache, runtime overhead included.

GPUs and Macs that run Gemma 4 12B

The best way Gemma 4 12B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40603-bit96K31 tok/s1.6 sPDF: Gemma 4 12B on RTX 4060 report
RTX 50603-bit96K50 tok/s1.3 sPDF: Gemma 4 12B on RTX 5060 report
RTX 30703-bit96K50 tok/s1.2 sPDF: Gemma 4 12B on RTX 3070 report
RTX 3080 10 GB5-bit24K57 tok/s0.8 sPDF: Gemma 4 12B on RTX 3080 10 GB report
RTX 3060 12 GB6-bit112K25 tok/s1.9 sPDF: Gemma 4 12B on RTX 3060 12 GB report
Arc B5806-bit112K22 tok/s1.3 sPDF: Gemma 4 12B on Arc B580 report
RTX 40706-bit112K34 tok/s0.8 sPDF: Gemma 4 12B on RTX 4070 report
RTX 4070 Super6-bit112K34 tok/s0.7 sPDF: Gemma 4 12B on RTX 4070 Super report
RTX 50706-bit112K45 tok/s0.8 sPDF: Gemma 4 12B on RTX 5070 report
RTX 4060 Ti 16 GB8-bit248K16 tok/s1.1 sPDF: Gemma 4 12B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit248K24 tok/s1.0 sPDF: Gemma 4 12B on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit248K21 tok/s1.2 sPDF: Gemma 4 12B on Arc A770 16 GB report
Radeon RX 9070 XT8-bit248K28 tok/s0.7 sPDF: Gemma 4 12B on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit248K35 tok/s0.6 sPDF: Gemma 4 12B on RTX 4070 Ti Super report
RTX 4080 Super8-bit248K38 tok/s0.5 sPDF: Gemma 4 12B on RTX 4080 Super report
RTX 5070 Ti8-bit248K46 tok/s0.6 sPDF: Gemma 4 12B on RTX 5070 Ti report
RTX 50808-bit248K49 tok/s0.4 sPDF: Gemma 4 12B on RTX 5080 report
Radeon RX 7900 XTX8-bit256K41 tok/s1.1 sPDF: Gemma 4 12B on Radeon RX 7900 XTX report
RTX 30908-bit256K48 tok/s0.7 sPDF: Gemma 4 12B on RTX 3090 report
RTX 40908-bit256K51 tok/s0.3 sPDF: Gemma 4 12B on RTX 4090 report
Radeon AI PRO R97008-bit256K28 tok/s0.7 sPDF: Gemma 4 12B on Radeon AI PRO R9700 report
RTX 50908-bit256K86 tok/s0.2 sPDF: Gemma 4 12B on RTX 5090 report
2× RTX 30908-bit256K48 tok/s0.4 sPDF: Gemma 4 12B on 2× RTX 3090 report
2× RTX 40908-bit256K51 tok/s0.2 sPDF: Gemma 4 12B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit256K67 tok/s0.4 sPDF: Gemma 4 12B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit256K49 tok/s0.3 sPDF: Gemma 4 12B on RTX 6000 Ada report
L40S8-bit256K45 tok/s0.4 sPDF: Gemma 4 12B on L40S report
2× RTX 50908-bit256K86 tok/s0.1 sPDF: Gemma 4 12B on 2× RTX 5090 report
4× RTX 30908-bit256K48 tok/s0.2 sPDF: Gemma 4 12B on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit256K86 tok/s0.2 sPDF: Gemma 4 12B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB5-bit128K11 tok/s16 sPDF: Gemma 4 12B on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit256K7.5 tok/s16 sPDF: Gemma 4 12B on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit256K7.5 tok/s16 sPDF: Gemma 4 12B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit256K17 tok/s7.8 sPDF: Gemma 4 12B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit256K17 tok/s7.8 sPDF: Gemma 4 12B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit256K17 tok/s7.8 sPDF: Gemma 4 12B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit256K29 tok/s3.9 sPDF: Gemma 4 12B on M4 Max 128 GB report
M5 Max 128 GB8-bit256K32 tok/s1.1 sPDF: Gemma 4 12B on M5 Max 128 GB report
M3 Ultra 96 GB8-bit256K36 tok/s2.2 sPDF: Gemma 4 12B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit256K36 tok/s2.2 sPDF: Gemma 4 12B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit256K36 tok/s2.2 sPDF: Gemma 4 12B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit256K50 tok/s0.7 sPDF: Gemma 4 12B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit256K12 tok/s2.2 sPDF: Gemma 4 12B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit256K12 tok/s2.2 sPDF: Gemma 4 12B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit256K16 tok/s0.8 sPDF: Gemma 4 12B on RTX Spark report
DGX Spark8-bit256K15 tok/s0.8 sPDF: Gemma 4 12B on DGX Spark report
Dual DGX Spark8-bit256K15 tok/s0.4 sPDF: Gemma 4 12B on Dual DGX Spark report
DGX Station GB3008-bit256K250 tok/s0.1 sPDF: Gemma 4 12B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit256K86 tok/s0.1 sPDF: Gemma 4 12B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit256K86 tok/s0.1 sPDF: Gemma 4 12B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit256K86 tok/s0.1 sPDF: Gemma 4 12B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit256K97 tok/s0.5 sPDF: Gemma 4 12B on A100 80 GB SXM report
H100 SXM8-bit256K146 tok/s0.2 sPDF: Gemma 4 12B on H100 SXM report
4× H200 NVL8-bit256K191 tok/s0.1 sPDF: Gemma 4 12B on 4× H200 NVL report
8× H200 NVL8-bit256K191 tok/s0.1 sPDF: Gemma 4 12B on 8× H200 NVL report

Best fit is the highest quality, up to 8-bit, that leaves 16K of context. First reply is the wait for a 2K prompt. Speeds are rough, plus or minus 25%.

How to run it

Ollama
ollama run gemma4:12b
llama.cpp
llama-server -hf unsloth/gemma-4-12b-it-GGUF
vLLM
vllm serve google/gemma-4-12B-it
MLX (Mac)
mlx_lm.chat --model mlx-community/gemma-4-12B-it-4bit

Questions

How much VRAM does Gemma 4 12B need?

Gemma 4 12B needs 6.8 GB for its weights at 4-bit, 8.2 GB with an 8K context and 8.4 GB with 32K. At 8-bit, 13.3 GB.

What is the minimum VRAM for Gemma 4 12B?

About 6.9 GB at 3-bit with an 8K context, the lowest precision we recommend for a model this size. The smallest machine here that runs it without spilling is the RTX 4060.

How big is Gemma 4 12B?

12B parameters: 22.3 GB of weights at 16-bit, 6.8 GB at 4-bit.

Can Gemma 4 12B run on an RTX 4090?

Yes. At 8-bit it holds 256K of context and writes about 51 tokens a second.

More models