Back to search results
Llama-3.2-3B-Instruct · IQ4_XS
NVIDIA RTX 507012GBllama.cpp v0.40L2 Cross-framework verified
This page aggregates 1 real-world runs of Llama-3.2-3B-Instruct (IQ4_XS) on NVIDIA RTX 5070 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.
238.16 tok/s
Decode
Decode speed
793.59 tok/s
Prefill
Prefill speed
0.6 s
TTFT
Time to first token
6.76 GB
VRAM
VRAM usage
L2 Cross-framework verified
1
Measured runs
1
Independent sources
Other
Source platforms
Today
Last verified
Performance
- llama.cpp · IQ4_XS (current)Decode 238.16 · Prefill 793.59 ·
- llama.cpp · Q5_K_S Decode 193.78 · Prefill 821.91 ·
- Ollama · FP16 Decode 100.33 · Prefill 407.38 ·
- Ollama · Q3_K_L Decode 203.4 · Prefill 838.81 ·
- llama.cpp · Q8_0 Decode 163.8 · Prefill 1228.5 ·
- llama.cpp · Q5_K_M Decode 238.3 · Prefill 1787.3 ·
- llama.cpp · Q4_K_M Decode 291.2 · Prefill 2184 ·
Core figures
- Decode (avg)
- 238.16 tok/s
- Prefill (avg)
- 793.59 tok/s
- TTFT (avg)
- 0.6 s
- VRAM (avg)
- 6.76 GB
- MTP acceptance rate
- —
- TTFB
- 1.91 GB
- Power draw
- 204.7 W
Configuration
Member-level fields are taken from the most recent run
- Model
- Llama-3.2-3B-Instruct
- Quantization
- IQ4_XS
- Framework
- llama.cpp
- Version
- v0.40
- Context length
- 32768 tokens
- Batch size
- 512
- GPU layers
- 28
- Flash Attention
- On
Hardware
Nominal and measured figures are shown side by side; whether it runs is the reader's call
- GPU
- NVIDIA RTX 5070
- Nominal VRAM
- 12 GB
- Measured VRAM (avg)
- 6.76 GB
- OS
- Windows 11 24H2
- Driver
- 572.83
- CUDA
- 12.6
- Power draw
- 204.7 W
Sources & evidence
1 measured records in total, each traceable to its original source
v0.40 · Windows 11 24H2 · CUDA 12.6 · 32768 ctx
238.16 tok/s
Decode
793.59 tok/s
Prefill
0.6 s
TTFT
6.76 GB
VRAM
—
MTP
204.7 W
Power draw
L2 batch seed 2026-10-18: llama.cpp on hw_id=18 running model_id=70 at IQ4_XS