Llama-3-8B-Instruct · Q4_K_M
This page aggregates 2 real-world runs of Llama-3-8B-Instruct (Q4_K_M) on NVIDIA RTX 4090 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.
146.97 tok/s
Decode
Decode speed
902.58 tok/s
Prefill
Prefill speed
0.68 s
TTFT
Time to first token
10.64 GB
VRAM
VRAM usage
2
Measured runs
1
Independent sources
Other
Source platforms
2026-10-18
Last verified
Performance
- llama.cpp · Q5_K_S Decode 118.71 · Prefill 394.1 ·
- vLLM · IQ3_S Decode 177.41 · Prefill 791.84 ·
- vLLM · Q4_K_S Decode 157.4 · Prefill 646.64 ·
- llama.cpp · Q4_K_M (current)Decode 146.97 · Prefill 902.58 ·
- llama.cpp · Q4_K_S Decode 134.4 · Prefill 594.59 ·
- Strata · Q2_K Decode 254.93 · Prefill 1284 ·
- llama.cpp · Q8_0 Decode 91.4 · Prefill 685.5 ·
- llama.cpp · Q5_K_M Decode 133.3 · Prefill 999.8 ·
Core figures
- Decode (avg)
- 146.97 tok/s
- Prefill (avg)
- 902.58 tok/s
- TTFT (avg)
- 0.68 s
- VRAM (avg)
- 10.64 GB
- MTP acceptance rate
- —
- TTFB
- 1.07 GB
- Power draw
- 358.8 W
Configuration
Member-level fields are taken from the most recent run
- Model
- Llama-3-8B-Instruct
- Quantization
- Q4_K_M
- Framework
- llama.cpp
- Version
- v0.27
- Context length
- 8192 tokens
- Batch size
- 512
- GPU layers
- 99
- Flash Attention
- Off
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)
- 10.64 GB
- OS
- Windows 11 23H2
- Driver
- 560.94
- CUDA
- 13.0
- Power draw
- 358.8 W
Sources & evidence
2 measured records in total, each traceable to its original source
v0.27 · Windows 11 23H2 · CUDA 13.0 · 8192 ctx
133.43 tok/s
Decode
601.35 tok/s
Prefill
1.28 s
TTFT
15.88 GB
VRAM
—
MTP
358.8 W
Power draw
L2 batch seed 2026-10-18: llama.cpp on hw_id=1 running model_id=69 at Q4_K_M
b4200 · Linux · CUDA 12.6 · 4096 ctx
160.5 tok/s
Decode
1,203.8 tok/s
Prefill
0.08 s
TTFT
5.4 GB
VRAM
—
MTP
— W
Power draw
L0 理论估算(add-1000-records-batch3)