Arao AI
Back to search results

Qwen3.5-27B · IQ3_S

Start a discussion
Apple M4 Max128GBllama.cpp v0.45L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3.5-27B (IQ3_S) on Apple M4 Max with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

165.6 tok/s

Decode

Decode speed

722.22 tok/s

Prefill

Prefill speed

1.81 s

TTFT

Time to first token

39.52 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

Today

Last verified

Performance

  1. llama.cpp · IQ3_S (current)Decode 165.6 · Prefill 722.22 ·
  2. vLLM · Q3_K_M Decode 166.04 · Prefill 735.49 ·
  3. Ollama · Q3_K_S Decode 165.47 · Prefill 649.98 ·
  4. Ollama · Q5_K_S Decode 111.59 · Prefill 447.41 ·

Core figures

Decode (avg)
165.6 tok/s
Prefill (avg)
722.22 tok/s
TTFT (avg)
1.81 s
VRAM (avg)
39.52 GB
MTP acceptance rate
—
TTFB
1.32 GB
Power draw
73.7 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3.5-27B
Quantization
IQ3_S
Framework
llama.cpp
Version
v0.45
Context length
2048 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 M4 Max
Nominal VRAM
128 GB
Measured VRAM (avg)
39.52 GB
OS
macOS 15.0
Driver
Apple M4 Max
CUDA
N/A
Power draw
73.7 W

Sources & evidence

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

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

    v0.45 · macOS 15.0 · CUDA N/A · 2048 ctx

    165.6 tok/s

    Decode

    722.22 tok/s

    Prefill

    1.81 s

    TTFT

    39.52 GB

    VRAM

    —

    MTP

    73.7 W

    Power draw

    L2 batch seed 2026-10-10: llama.cpp on hw_id=26 running model_id=50 at IQ3_S