Back to search results
Qwen3-4B · IQ4_XS
NVIDIA RTX 509032GBllama.cpp v0.42L2 Cross-framework verified
This page aggregates 1 real-world runs of Qwen3-4B (IQ4_XS) on NVIDIA RTX 5090 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.
335.8 tok/s
Decode
Decode speed
1,082.78 tok/s
Prefill
Prefill speed
2.49 s
TTFT
Time to first token
8.32 GB
VRAM
VRAM usage
L2 Cross-framework verified
1
Measured runs
1
Independent sources
Other
Source platforms
2 days ago
Last verified
Performance
- llama.cpp · FP16 Decode 149.3 · Prefill 589.87 ·
- llama.cpp · IQ4_XS (current)Decode 335.8 · Prefill 1082.78 ·
- Ollama · IQ3_XS Decode 349.78 · Prefill 1805.86 ·
Core figures
- Decode (avg)
- 335.8 tok/s
- Prefill (avg)
- 1,082.78 tok/s
- TTFT (avg)
- 2.49 s
- VRAM (avg)
- 8.32 GB
- MTP acceptance rate
- —
- TTFB
- 2.69 GB
- Power draw
- 283.7 W
Configuration
Member-level fields are taken from the most recent run
- Model
- Qwen3-4B
- Quantization
- IQ4_XS
- Framework
- llama.cpp
- Version
- v0.42
- Context length
- 4096 tokens
- Batch size
- 256
- 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)
- 8.32 GB
- OS
- macOS 15.0
- Driver
- 560.94
- CUDA
- 12.6
- Power draw
- 283.7 W
Sources & evidence
1 measured records in total, each traceable to its original source
v0.42 · macOS 15.0 · CUDA 12.6 · 4096 ctx
335.8 tok/s
Decode
1,082.78 tok/s
Prefill
2.49 s
TTFT
8.32 GB
VRAM
—
MTP
283.7 W
Power draw
L2 batch seed 2026-10-16: llama.cpp on hw_id=13 running model_id=54 at IQ4_XS