Back to search results

gemma-4-12B-it · Q8_0

Start a discussion
NVIDIA RTX 509032GBllama.cpp v0.21L2 Cross-framework verified

This page aggregates 2 real-world runs of gemma-4-12B-it (Q8_0) on NVIDIA RTX 5090 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

125.52 tok/s

Decode

Decode speed

751.15 tok/s

Prefill

Prefill speed

0.88 s

TTFT

Time to first token

21.95 GB

VRAM

VRAM usage

L2 Cross-framework verified

2

Measured runs

1

Independent sources

Other

Source platforms

2026-10-19

Last verified

Performance

  1. llama.cpp · IQ2_XS Decode 262.05 · Prefill 997.54 ·
  2. llama.cpp · Q8_0 (current)Decode 125.52 · Prefill 751.15 ·
  3. Ollama · Q3_K_S Decode 236.13 · Prefill 819.72 ·
  4. Ollama · Q5_K_S Decode 165.12 · Prefill 690.64 ·
  5. vLLM · IQ3_S Decode 261.31 · Prefill 1085.12 ·
  6. vLLM · Q5_K_S Decode 173.67 · Prefill 792.94 ·
  7. Strata · Q3_K_M Decode 340.87 · Prefill 1525.66 ·
  8. llama.cpp · IQ3_S Decode 214.27 · Prefill 831.85 ·
  9. vLLM · IQ1_S Decode 298.32 · Prefill 1152.98 ·
  10. vLLM · IQ2_M Decode 288.51 · Prefill 940.03 ·
  11. Ollama · IQ2_XS Decode 246.6 · Prefill 1083.78 ·
  12. vLLM · Q8_0 Decode 117.2 · Prefill 879 ·
  13. vLLM · Q5_K_M Decode 171.1 · Prefill 1283.3 ·
  14. vLLM · Q4_K_M Decode 208.7 · Prefill 1565.3 ·
  15. llama.cpp · Q5_K_M Decode 157.1 · Prefill 1178.3 ·
  16. llama.cpp · Q4_K_M Decode 191.5 · Prefill 1436.3 ·

Core figures

Decode (avg)
125.52 tok/s
Prefill (avg)
751.15 tok/s
TTFT (avg)
0.88 s
VRAM (avg)
21.95 GB
MTP acceptance rate
—
TTFB
1.15 GB
Power draw
388.6 W

Configuration

Member-level fields are taken from the most recent run

Model
gemma-4-12B-it
Quantization
Q8_0
Framework
llama.cpp
Version
v0.21
Context length
16384 tokens
Batch size
512
GPU layers
33
Flash Attention
On

Hardware

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

GPU
NVIDIA RTX 5090
Nominal VRAM
32 GB
Measured VRAM (avg)
21.95 GB
OS
Windows 11 24H2
Driver
570.86
CUDA
12.6
Power draw
388.6 W

Sources & evidence

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

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

    v0.21 · Windows 11 24H2 · CUDA 12.6 · 16384 ctx

    143.53 tok/s

    Decode

    695.99 tok/s

    Prefill

    1.65 s

    TTFT

    30.4 GB

    VRAM

    —

    MTP

    388.6 W

    Power draw

    L2 batch seed 2026-10-19: llama.cpp on hw_id=13 running model_id=65 at Q8_0
  2. L0 Self-reportedOtherOriginal link Verified on 2026-10-06

    b4200 · Linux · CUDA 12.6 · 4096 ctx

    107.5 tok/s

    Decode

    806.3 tok/s

    Prefill

    0.12 s

    TTFT

    13.5 GB

    VRAM

    —

    MTP

    — W

    Power draw

    L0 理论估算(add-1000-records-batch3)