Arao AI
Back to search results

gemma-4-12B-it · Q3_K_L

Start a discussion
NVIDIA RTX 409024GBllama.cpp v0.23L2 Cross-framework verified

This page aggregates 1 real-world runs of gemma-4-12B-it (Q3_K_L) on NVIDIA RTX 4090 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

114.65 tok/s

Decode

Decode speed

473.3 tok/s

Prefill

Prefill speed

2.16 s

TTFT

Time to first token

20.24 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

16 days ago

Last verified

Performance

  1. llama.cpp · Q3_K_L (current)Decode 114.65 · Prefill 473.3 ·
  2. Ollama · Q4_K_M Decode 87.43 · Prefill 399.69 ·
  3. Ollama · Q6_K Decode 70.15 · Prefill 334.66 ·
  4. llama.cpp · IQ2_XS Decode 152.32 · Prefill 683.44 ·
  5. Strata · Q5_K_S Decode 117.55 · Prefill 384.33 ·
  6. llama.cpp · IQ3_XS Decode 114.16 · Prefill 511.31 ·
  7. llama.cpp · Q4_K_S Decode 100.24 · Prefill 360.86 ·
  8. vLLM · Q2_K Decode 139.41 · Prefill 531.29 ·
  9. Ollama · IQ4_XS Decode 95.26 · Prefill 447.14 ·
  10. vLLM · Q8_0 Decode 65.9 · Prefill 494.3 ·
  11. vLLM · Q5_K_M Decode 96.3 · Prefill 722.3 ·
  12. vLLM · Q4_K_M Decode 117.4 · Prefill 880.5 ·
  13. llama.cpp · Q8_0 Decode 60.5 · Prefill 453.8 ·
  14. llama.cpp · Q5_K_M Decode 88.3 · Prefill 662.3 ·
  15. llama.cpp · Q4_K_M Decode 107.7 · Prefill 807.8 ·

Core figures

Decode (avg)
114.65 tok/s
Prefill (avg)
473.3 tok/s
TTFT (avg)
2.16 s
VRAM (avg)
20.24 GB
MTP acceptance rate
—
TTFB
0.92 GB
Power draw
348.8 W

Configuration

Member-level fields are taken from the most recent run

Model
gemma-4-12B-it
Quantization
Q3_K_L
Framework
llama.cpp
Version
v0.23
Context length
2048 tokens
Batch size
512
GPU layers
20
Flash Attention
On

Hardware

Nominal and measured figures are shown side by side; whether it runs is the reader's call

GPU
NVIDIA RTX 4090
Nominal VRAM
24 GB
Measured VRAM (avg)
20.24 GB
OS
Ubuntu 24.04
Driver
565.57
CUDA
12.4
Power draw
348.8 W

Sources & evidence

1 measured records in total, each traceable to its original source

  1. L2 Cross-framework verifiedOtherOriginal link Verified on 2026-10-19

    v0.23 · Ubuntu 24.04 · CUDA 12.4 · 2048 ctx

    114.65 tok/s

    Decode

    473.3 tok/s

    Prefill

    2.16 s

    TTFT

    20.24 GB

    VRAM

    —

    MTP

    348.8 W

    Power draw

    L2 batch seed 2026-10-19: llama.cpp on hw_id=1 running model_id=65 at Q3_K_L