Arao AI
Back to search results

Qwen3.5-4B · IQ3_S

Start a discussion
Apple M5 Max128GBllama.cpp v0.25L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3.5-4B (IQ3_S) on Apple M5 Max with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

279.98 tok/s

Decode

Decode speed

1,354 tok/s

Prefill

Prefill speed

0.12 s

TTFT

Time to first token

7.64 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 · IQ3_S (current)Decode 279.98 · Prefill 1354 ·
  2. Strata · Q3_K_L Decode 354.31 · Prefill 1310.91 ·
  3. Strata · Q5_K_S Decode 279.1 · Prefill 1192.93 ·
  4. vLLM · IQ2_M Decode 364.07 · Prefill 1250.03 ·
  5. vLLM · Q4_K_S Decode 256.46 · Prefill 848.78 ·
  6. Strata · Q5_K_M Decode 272.64 · Prefill 963.9 ·

Core figures

Decode (avg)
279.98 tok/s
Prefill (avg)
1,354 tok/s
TTFT (avg)
0.12 s
VRAM (avg)
7.64 GB
MTP acceptance rate
—
TTFB
2.24 GB
Power draw
164.3 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3.5-4B
Quantization
IQ3_S
Framework
llama.cpp
Version
v0.25
Context length
16384 tokens
Batch size
512
GPU layers
99
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)
7.64 GB
OS
Ubuntu 22.04
Driver
555.85
CUDA
12.4
Power draw
164.3 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

    v0.25 · Ubuntu 22.04 · CUDA 12.4 · 16384 ctx

    279.98 tok/s

    Decode

    1,354 tok/s

    Prefill

    0.12 s

    TTFT

    7.64 GB

    VRAM

    —

    MTP

    164.3 W

    Power draw

    L2 batch seed 2026-10-16: llama.cpp on hw_id=28 running model_id=48 at IQ3_S