gemma-4-12B-it · Q4_K_M
This page aggregates 2 real-world runs of gemma-4-12B-it (Q4_K_M) on NVIDIA RTX 3090 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.
79.26 tok/s
Decode
Decode speed
473.41 tok/s
Prefill
Prefill speed
1.21 s
TTFT
Time to first token
15.28 GB
VRAM
VRAM usage
2
Measured runs
1
Independent sources
Other
Source platforms
2026-10-19
Last verified
Performance
- llama.cpp · Q4_K_M (current)Decode 79.26 · Prefill 473.41 ·
- Ollama · Q3_K_L Decode 64.09 · Prefill 302.97 ·
- vLLM · Q6_K Decode 50.6 · Prefill 208.67 ·
- Ollama · IQ3_XS Decode 67.34 · Prefill 278.97 ·
- vLLM · Q4_K_M Decode 100.66 · Prefill 612.88 ·
- Ollama · Q2_K Decode 115.28 · Prefill 513.9 ·
- Ollama · Q4_K_S Decode 84.3 · Prefill 357.49 ·
- Strata · IQ2_M Decode 174.12 · Prefill 755.26 ·
- vLLM · Q8_0 Decode 61.2 · Prefill 459 ·
- vLLM · Q5_K_M Decode 89.4 · Prefill 670.5 ·
- llama.cpp · Q8_0 Decode 56.2 · Prefill 421.5 ·
- llama.cpp · Q5_K_M Decode 82 · Prefill 615 ·
Core figures
- Decode (avg)
- 79.26 tok/s
- Prefill (avg)
- 473.41 tok/s
- TTFT (avg)
- 1.21 s
- VRAM (avg)
- 15.28 GB
- MTP acceptance rate
- —
- TTFB
- 0.47 GB
- Power draw
- 282.1 W
Configuration
Member-level fields are taken from the most recent run
- Model
- gemma-4-12B-it
- Quantization
- Q4_K_M
- Framework
- llama.cpp
- Version
- v0.50
- Context length
- 16384 tokens
- Batch size
- 512
- GPU layers
- 99
- Flash Attention
- On
Hardware
Nominal and measured figures are shown side by side; whether it runs is the reader's call
- GPU
- NVIDIA RTX 3090
- Nominal VRAM
- 24 GB
- Measured VRAM (avg)
- 15.28 GB
- OS
- Ubuntu 22.04
- Driver
- 560.94
- CUDA
- 12.6
- Power draw
- 282.1 W
Sources & evidence
2 measured records in total, each traceable to its original source
v0.50 · Ubuntu 22.04 · CUDA 12.6 · 16384 ctx
58.52 tok/s
Decode
196.82 tok/s
Prefill
2.3 s
TTFT
22.76 GB
VRAM
—
MTP
282.1 W
Power draw
L2 batch seed 2026-10-19: llama.cpp on hw_id=2 running model_id=65 at Q4_K_M
b4200 · Linux · CUDA 12.6 · 4096 ctx
100 tok/s
Decode
750 tok/s
Prefill
0.13 s
TTFT
7.8 GB
VRAM
—
MTP
— W
Power draw
L0 理论估算(add-1000-records-batch3)