Back to search results
Llama-3.2-3B-Instruct · FP16
NVIDIA RTX 309024GBvLLM v0.26L2 Cross-framework verified
This page aggregates 1 real-world runs of Llama-3.2-3B-Instruct (FP16) on NVIDIA RTX 3090 with vLLM, contributed by 1 independent source platforms; metrics are averages of published measurements.
113.95 tok/s
Decode
Decode speed
381.37 tok/s
Prefill
Prefill speed
1.4 s
TTFT
Time to first token
13.76 GB
VRAM
VRAM usage
L2 Cross-framework verified
1
Measured runs
1
Independent sources
Other
Source platforms
16 days ago
Last verified
Performance
- llama.cpp · Q2_K Decode 287.44 · Prefill 1386.85 ·
- vLLM · FP16 (current)Decode 113.95 · Prefill 381.37 ·
- vLLM · Q3_K_M Decode 250.47 · Prefill 1151.95 ·
- Strata · FP16 Decode 145.3 · Prefill 693.51 ·
- llama.cpp · Q8_0 Decode 228.2 · Prefill 1711.5 ·
- llama.cpp · Q5_K_M Decode 331.9 · Prefill 2489.3 ·
- llama.cpp · Q4_K_M Decode 405.6 · Prefill 3042 ·
Core figures
- Decode (avg)
- 113.95 tok/s
- Prefill (avg)
- 381.37 tok/s
- TTFT (avg)
- 1.4 s
- VRAM (avg)
- 13.76 GB
- MTP acceptance rate
- —
- TTFB
- 0.91 GB
- Power draw
- 288.3 W
Configuration
Member-level fields are taken from the most recent run
- Model
- Llama-3.2-3B-Instruct
- Quantization
- FP16
- Framework
- vLLM
- Version
- v0.26
- Context length
- 32768 tokens
- Batch size
- 8
- GPU layers
- 999
- Flash Attention
- Off
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)
- 13.76 GB
- OS
- Windows 11 24H2
- Driver
- 565.57
- CUDA
- 12.8
- Power draw
- 288.3 W
Sources & evidence
1 measured records in total, each traceable to its original source
v0.26 · Windows 11 24H2 · CUDA 12.8 · 32768 ctx
113.95 tok/s
Decode
381.37 tok/s
Prefill
1.4 s
TTFT
13.76 GB
VRAM
—
MTP
288.3 W
Power draw
L2 batch seed 2026-10-19: vLLM on hw_id=2 running model_id=70 at FP16