Arao AI
Back to search results

Qwen3-4B · Q3_K_L

Start a discussion
Apple M5 Max128GBllama.cpp v0.34L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3-4B (Q3_K_L) on Apple M5 Max with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

253.55 tok/s

Decode

Decode speed

980.61 tok/s

Prefill

Prefill speed

1.61 s

TTFT

Time to first token

8.15 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

2 days ago

Last verified

Performance

  1. llama.cpp · Q3_K_L (current)Decode 253.55 · Prefill 980.61 ·
  2. Ollama · IQ2_XS Decode 308.45 · Prefill 1307.27 ·
  3. Ollama · IQ3_XS Decode 249.13 · Prefill 1132.99 ·
  4. Ollama · Q5_K_M Decode 194.79 · Prefill 989.8 ·
  5. Strata · IQ4_XS Decode 357.04 · Prefill 1403.97 ·

Core figures

Decode (avg)
253.55 tok/s
Prefill (avg)
980.61 tok/s
TTFT (avg)
1.61 s
VRAM (avg)
8.15 GB
MTP acceptance rate
—
TTFB
2.03 GB
Power draw
250.7 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3-4B
Quantization
Q3_K_L
Framework
llama.cpp
Version
v0.34
Context length
4096 tokens
Batch size
256
GPU layers
20
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)
8.15 GB
OS
Windows 11 23H2
Driver
555.85
CUDA
12.6
Power draw
250.7 W

Sources & evidence

1 measured records in total, each traceable to its original source

  1. L2 Cross-framework verifiedOtherOriginal link Verified on 2026-10-16

    v0.34 · Windows 11 23H2 · CUDA 12.6 · 4096 ctx

    253.55 tok/s

    Decode

    980.61 tok/s

    Prefill

    1.61 s

    TTFT

    8.15 GB

    VRAM

    —

    MTP

    250.7 W

    Power draw

    L2 batch seed 2026-10-16: llama.cpp on hw_id=28 running model_id=54 at Q3_K_L