dexiio

What hardware runs Gemma 4 12B?

Gemma 4 12B is a 11.91B-parameter dense model (apache-2.0 license). June 2026's 16GB-class headline: dense 12B with native vision and audio in one backbone. The new laptop ceiling.

All figures assume an f16 KV cache, a 0.6 GB display reserve on the GPU, and 64 GB of DDR5 system RAM for the offload tiers. Tune these in the calculator.

Minimum VRAM by quant and context

QuantFile size8K16K32K64K128K
Q3_K_M *5.3 GB8 GB8.7 GB9.9 GB12.4 GB17.2 GB
Q4_K_M6.6 GB9.4 GB10 GB11.2 GB13.7 GB18.5 GB
Q5_K_M7.8 GB10.6 GB11.2 GB12.4 GB14.9 GB19.7 GB
Q6_K9.1 GB11.9 GB12.5 GB13.7 GB16.2 GB21 GB
Q8_011.8 GB14.5 GB15.2 GB16.4 GB18.9 GB23.7 GB
IQ4_XS5.9 GB8.7 GB9.3 GB10.6 GB13.1 GB17.8 GB

Full-GPU figures: weights + f16 KV cache + overhead. * below our recommended floor of Q4_K_M.

GPU compatibility

GPU8K16K32K64K128K
NVIDIA GeForce RTX 5090Q8_0Q8_0Q8_0Q8_0Q8_0
NVIDIA GeForce RTX 5080Q8_0Q8_0Q6_KQ5_K_MIQ4_XS
NVIDIA GeForce RTX 5070 TiQ8_0Q8_0Q6_KQ5_K_MIQ4_XS
NVIDIA GeForce RTX 5070Q6_KQ5_K_MQ4_K_MIQ4_XSQ8_0
NVIDIA GeForce RTX 5060 Ti 16GBQ8_0Q8_0Q6_KQ5_K_MIQ4_XS
NVIDIA GeForce RTX 5060IQ4_XSIQ4_XSIQ4_XSIQ4_XSQ8_0
NVIDIA GeForce RTX 4090Q8_0Q8_0Q8_0Q8_0Q8_0
NVIDIA GeForce RTX 4080 SUPERQ8_0Q8_0Q6_KQ5_K_MIQ4_XS
NVIDIA GeForce RTX 4070 Ti SUPERQ8_0Q8_0Q6_KQ5_K_MIQ4_XS
NVIDIA GeForce RTX 4070Q6_KQ5_K_MQ4_K_MIQ4_XSQ8_0
NVIDIA GeForce RTX 4060 Ti 16GBQ8_0Q8_0Q6_KQ5_K_MIQ4_XS
NVIDIA GeForce RTX 4060IQ4_XSIQ4_XSIQ4_XSIQ4_XSQ8_0
NVIDIA GeForce RTX 3090Q8_0Q8_0Q8_0Q8_0Q8_0
NVIDIA GeForce RTX 3080 10GBQ4_K_MQ4_K_MIQ4_XSIQ4_XSQ8_0
NVIDIA GeForce RTX 3070IQ4_XSIQ4_XSIQ4_XSIQ4_XSQ8_0
NVIDIA GeForce RTX 3060 TiIQ4_XSIQ4_XSIQ4_XSIQ4_XSQ8_0
NVIDIA GeForce RTX 3060 12GBQ6_KQ5_K_MQ4_K_MIQ4_XSQ8_0
NVIDIA GeForce RTX 2080 TiQ5_K_MQ4_K_MIQ4_XSIQ4_XSQ8_0
NVIDIA GeForce GTX 1080 TiQ5_K_MQ4_K_MIQ4_XSIQ4_XSQ8_0
AMD Radeon RX 9070 XTQ8_0Q8_0Q6_KQ5_K_MIQ4_XS
AMD Radeon RX 7900 XTXQ8_0Q8_0Q8_0Q8_0Q8_0
AMD Radeon RX 7900 XTQ8_0Q8_0Q8_0Q8_0Q5_K_M
AMD Radeon RX 7800 XTQ8_0Q8_0Q6_KQ5_K_MIQ4_XS
AMD Radeon RX 7600 XTQ8_0Q8_0Q6_KQ5_K_MIQ4_XS
AMD Radeon RX 6800 XTQ8_0Q8_0Q6_KQ5_K_MIQ4_XS
AMD Radeon RX 6700 XTQ6_KQ5_K_MQ4_K_MIQ4_XSQ8_0
Intel Arc B580Q6_KQ5_K_MQ4_K_MIQ4_XSQ8_0
Intel Arc A770 16GBQ8_0Q8_0Q6_KQ5_K_MIQ4_XS
Apple M2 (16GB)Q6_KQ5_K_MQ4_K_MIQ4_XSQ8_0
Apple M2 Pro (32GB)Q8_0Q8_0Q8_0Q8_0Q8_0
Apple M2 Max (64GB)Q8_0Q8_0Q8_0Q8_0Q8_0
Apple M2 Ultra (128GB)Q8_0Q8_0Q8_0Q8_0Q8_0
Apple M3 (16GB)Q6_KQ5_K_MQ4_K_MIQ4_XSQ8_0
Apple M3 Pro (36GB)Q8_0Q8_0Q8_0Q8_0Q8_0
Apple M3 Max (64GB)Q8_0Q8_0Q8_0Q8_0Q8_0
Apple M3 Ultra (96GB)Q8_0Q8_0Q8_0Q8_0Q8_0
Apple M4 (16GB)Q6_KQ5_K_MQ4_K_MIQ4_XSQ8_0
Apple M4 Pro (48GB)Q8_0Q8_0Q8_0Q8_0Q8_0
Apple M4 Max (64GB)Q8_0Q8_0Q8_0Q8_0Q8_0
CPU only / integrated graphicsQ8_0Q8_0Q8_0Q8_0Q8_0

Fits on GPUExpert offloadPartial offloadCPU only

Quant guidance

Our floor for Gemma 4 12B is Q4_K_M — below that, quality degrades faster than the VRAM savings are worth. Prefer the highest quant that still lands "Fits on GPU" at your context length in the table above.

Recommended run command

Q4_K_M at 32K on a NVIDIA GeForce RTX 2080 Ti-class GPU (Partial offload):

llama-server -m gemma-4-12b-it-Q4_K_M.gguf -c 32768 --flash-attn -ngl 46

Frequently asked questions

How much VRAM does Gemma 4 12B need?

At Q4_K_M and 32K context, Gemma 4 12B needs about 11.2 GB of VRAM to run fully on GPU (weights + KV cache + overhead).

What is the smallest GPU that can run Gemma 4 12B?

The NVIDIA GeForce RTX 2080 Ti (11 GB) is the smallest GPU in our set that runs Gemma 4 12B well at 32K context.

What quantization should I use for Gemma 4 12B?

We recommend Q4_K_M or higher. Q4_K_M weighs 6.6 GB (4.78 bits/weight); going below Q4_K_M costs noticeable quality on a model this size.

How long a context can Gemma 4 12B handle?

Gemma 4 12B supports up to 256K tokens. KV cache grows linearly with context, though sliding-window layers cap most of the growth.

Can I run Gemma 4 12B without a GPU?

Yes, at reduced speed: on CPU with 64 GB of DDR5 it manages roughly 4 tokens/sec at 8K context (Q8_0).

Related guides