Arao AI
Back to search results

Qwen3-8B · FP16

Start a discussion
Apple M4 Max128GBvLLM v0.34L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3-8B (FP16) on Apple M4 Max with vLLM, contributed by 1 independent source platforms; metrics are averages of published measurements.

103.62 tok/s

Decode

Decode speed

346.91 tok/s

Prefill

Prefill speed

1.64 s

TTFT

Time to first token

33.18 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

Today

Last verified

Performance

  1. llama.cpp · Q2_K Decode 238 · Prefill 1037.55 ·
  2. llama.cpp · Q3_K_L Decode 189.78 · Prefill 777.5 ·
  3. vLLM · FP16 (current)Decode 103.62 · Prefill 346.91 ·
  4. vLLM · IQ1_S Decode 287.05 · Prefill 1522.92 ·
  5. Ollama · FP16 Decode 81 · Prefill 424.01 ·
  6. Ollama · IQ4_XS Decode 182.06 · Prefill 996.81 ·

Core figures

Decode (avg)
103.62 tok/s
Prefill (avg)
346.91 tok/s
TTFT (avg)
1.64 s
VRAM (avg)
33.18 GB
MTP acceptance rate
—
TTFB
0.83 GB
Power draw
56.4 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3-8B
Quantization
FP16
Framework
vLLM
Version
v0.34
Context length
32768 tokens
Batch size
32
GPU layers
999
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)
33.18 GB
OS
macOS 15.0
Driver
Apple M4 Max
CUDA
N/A
Power draw
56.4 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

    v0.34 · macOS 15.0 · CUDA N/A · 32768 ctx

    103.62 tok/s

    Decode

    346.91 tok/s

    Prefill

    1.64 s

    TTFT

    33.18 GB

    VRAM

    —

    MTP

    56.4 W

    Power draw

    L2 batch seed 2026-10-10: vLLM on hw_id=26 running model_id=53 at FP16