Qwen2.5-Omni-3B · IQ3_S
This page aggregates 2 real-world runs of Qwen2.5-Omni-3B (IQ3_S) on Apple M5 Max with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.
341.31 tok/s
Decode
Decode speed
1,584.99 tok/s
Prefill
Prefill speed
1.01 s
TTFT
Time to first token
6.26 GB
VRAM
VRAM usage
2
Measured runs
1
Independent sources
Other
Source platforms
2 days ago
Last verified
Performance
- llama.cpp · IQ3_S (current)Decode 341.31 · Prefill 1584.99 ·
- llama.cpp · Q6_K Decode 221.16 · Prefill 1117.52 ·
- Strata · Q8_0 Decode 241.28 · Prefill 991.2 ·
- llama.cpp · IQ1_S Decode 289.86 · Prefill 1338.5 ·
Core figures
- Decode (avg)
- 341.31 tok/s
- Prefill (avg)
- 1,584.99 tok/s
- TTFT (avg)
- 1.01 s
- VRAM (avg)
- 6.26 GB
- MTP acceptance rate
- —
- TTFB
- 2.74 GB
- Power draw
- 159.1 W
Configuration
Member-level fields are taken from the most recent run
- Model
- Qwen2.5-Omni-3B
- Quantization
- IQ3_S
- Framework
- llama.cpp
- Version
- v0.31
- Context length
- 32768 tokens
- Batch size
- 256
- 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
- Apple M5 Max
- Nominal VRAM
- 128 GB
- Measured VRAM (avg)
- 6.26 GB
- OS
- Ubuntu 22.04
- Driver
- 560.94
- CUDA
- 12.4
- Power draw
- 207.7 W
Sources & evidence
2 measured records in total, each traceable to its original source
v0.31 · Ubuntu 22.04 · CUDA 12.4 · 32768 ctx
350.73 tok/s
Decode
1,896.04 tok/s
Prefill
0.75 s
TTFT
6.26 GB
VRAM
—
MTP
207.7 W
Power draw
L2 batch seed 2026-10-16: llama.cpp on hw_id=28 running model_id=60 at IQ3_S
v0.28 · macOS 15.0 · CUDA N/A · 16384 ctx
331.88 tok/s
Decode
1,273.94 tok/s
Prefill
1.28 s
TTFT
6.26 GB
VRAM
—
MTP
110.5 W
Power draw
L2 batch seed 2026-10-13: llama.cpp on hw_id=28 running model_id=60 at IQ3_S