Arao AI
Back to search results

Qwen3.8-9B · IQ3_XS

Start a discussion
Apple M5 Max128GBllama.cpp v0.41L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3.8-9B (IQ3_XS) on Apple M5 Max with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

79.08 tok/s

Decode

Decode speed

304.15 tok/s

Prefill

Prefill speed

1.65 s

TTFT

Time to first token

13.82 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

17 days ago

Last verified

Performance

  1. llama.cpp · IQ3_XS (current)Decode 79.08 · Prefill 304.15 ·
  2. Ollama · Q2_K Decode 78.89 · Prefill 425.19 ·
  3. Ollama · Q3_K_S Decode 73.12 · Prefill 238.27 ·
  4. vLLM · Q3_K_L Decode 202.52 · Prefill 1021.73 ·
  5. Strata · Q3_K_M Decode 280.05 · Prefill 1450.84 ·
  6. Strata · IQ3_XS Decode 320.14 · Prefill 1058.22 ·
  7. Strata · Q2_K Decode 331.24 · Prefill 1237.62 ·

Core figures

Decode (avg)
79.08 tok/s
Prefill (avg)
304.15 tok/s
TTFT (avg)
1.65 s
VRAM (avg)
13.82 GB
MTP acceptance rate
—
TTFB
0.63 GB
Power draw
60.1 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3.8-9B
Quantization
IQ3_XS
Framework
llama.cpp
Version
v0.41
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)
13.82 GB
OS
macOS 26.1
Power draw
60.1 W

Sources & evidence

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

  1. L2 Cross-framework verifiedOtherOriginal link Verified on 2026-10-19

    v0.41 · macOS 26.1 · 16384 ctx

    79.08 tok/s

    Decode

    304.15 tok/s

    Prefill

    1.65 s

    TTFT

    13.82 GB

    VRAM

    —

    MTP

    60.1 W

    Power draw

    L2 batch seed 2026-10-19: llama.cpp on hw_id=28 running model_id=52 at IQ3_XS