FitMyLLM
Google/Dense

GoogleGemma 4 12B

Gemma 4 12B — the mid-size dense model of the Gemma 4 generation. Vision-language, 256K context, and a 1024-token sliding window on five layers in six, which keeps its KV cache small for its size.

chatcodingreasoningmultilingualvisionmath
11.96B
Parameters
256K
Context length
16
Benchmarks
10
Quantizations
3.3M
HF downloads
Architecture
Dense
Released
2026-06-03
Layers
48
KV Heads
8
Head Dim
256
Family
gemma

Quantization Options

Context length:
QuantBitsVRAM @ 16KQuality
Q3_K_M4
7.0 GB
6.5 + 0.5 KV
low
Q3_K_L4.3
7.4 GB
6.9 + 0.5 KV
moderate
IQ4_XS4.46
7.7 GB
7.2 + 0.5 KV
moderate
Q4_K_S4.67
8.0 GB
7.5 + 0.5 KV
moderate
Q4_K_M4.89
8.3 GB
7.8 + 0.5 KV
good
Q5_K_S5.57
9.3 GB
8.8 + 0.5 KV
good
Q5_K_M5.7
9.5 GB
9.0 + 0.5 KV
good
Q6_K6.56
10.8 GB
10.3 + 0.5 KV
excellent
Q8_08.5
13.7 GB
13.2 + 0.5 KV
lossless
FP1616
24.9 GB
24.4 + 0.5 KV
lossless

Select your GPU above to see speed estimates and compatibility for each quantization.

Deploying for a team or in production? Size GPUs, cost & scaling in Enterprise →
Ready to run this?Rent by the hour

Rent a GPU and run Gemma 4 12B now

Spin up an A100 / H100 / 4090 in ~60s. Pay by the second. Cancel anytime.

Community Ratings

Loading ratings...

Benchmarks (16)

Run this model

Easiest way to get started·Beginners
Docs ↗
curl -fsSL https://ollama.com/install.sh | sh
$ollama run gemma4:12b-it-q4_K_M

Downloads and runs automatically. Add --verbose for speed stats.

Setup guide
>_

Auto-setup with fitmyllm CLI

Detects your GPU, recommends the best model, downloads it, and starts chatting — zero config. Benchmarks your speed and contributes anonymous data to improve predictions.

pip install fitmyllmthen run fitmyllmLearn more
Auto-detect GPULive tok/s in chatSpeed benchmarks9 inference engines

GPUs that can run this model

At Q4_K_M quantization. Sorted by minimum VRAM.

Apple M1 (8GB)
8 GB VRAM • 68 GB/s
APPLE
$499
Apple M2 (8GB)
8 GB VRAM • 100 GB/s
APPLE
$599
Apple M3 (8GB)
8 GB VRAM • 100 GB/s
APPLE
$599
NVIDIA Tesla K8
8 GB VRAM • 160 GB/s
NVIDIA
NVIDIA Tesla M60
8 GB VRAM • 160 GB/s
NVIDIA

Find the best GPU for Gemma 4 12B

Build Hardware for Gemma 4 12B
Spec sheet

Gemma 4 12B — 11.96B Dense.

Specifications
Parameters
11.96B
Architecture
Dense Transformer
Context Length
256K tokens
Capabilities
chat, coding, reasoning, multilingual, vision, math
Release Date
2026-06-03
Provider
Google
Family
gemma
VRAM requirements
QuantBPWVRAMQuality
Q3_K_M46.5 GB88%
Q3_K_L4.36.9 GB90%
IQ4_XS4.467.2 GB92%
Q4_K_S4.677.5 GB93%
Q4_K_M4.897.8 GB94%
Q5_K_S5.578.8 GB96%
Q5_K_M5.79.0 GB96%
Q6_K6.5610.3 GB97%
Q8_08.513.2 GB100%
FP161624.4 GB100%
Benchmark scores
MMLU-PRO77.2
GPQA Diamond75.3
LiveCodeBench72.0
AIME77.5
HLE15.7
AA Intelligence22.2
AA Coding31.0
IFBench73.5
Terminal-Bench Hard18.2
τ²-Bench36.3
SciCode38.2
AA Long Context61.7
MMMU Pro69.1
Terminal-Bench 2.127.3
EQ-Bench Creative Writing v31290
EQ-Bench 31416
Run command

Run Gemma 4 12B locally with Ollama — needs 7.8 GB VRAM at Q4_K_M:

$ollama run gemma4:12b
Compatible GPUs
30 @ Q4_K_M