Arao AI
Back to search results

Qwen3-4B-Thinking-2507 · FP16

Start a discussion
Apple M1 MaxvLLM v0.24L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3-4B-Thinking-2507 (FP16) on Apple M1 Max with vLLM, contributed by 1 independent source platforms; metrics are averages of published measurements.

48.27 tok/s

Decode

Decode speed

248.84 tok/s

Prefill

Prefill speed

0.2 s

TTFT

Time to first token

17.64 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

2 days ago

Last verified

Performance

  1. vLLM · FP16 (current)Decode 48.27 · Prefill 248.84 ·
  2. Strata · IQ3_XS Decode 182.49 · Prefill 683.44 ·
  3. Strata · Q4_K_M Decode 125.01 · Prefill 495.06 ·

Core figures

Decode (avg)
48.27 tok/s
Prefill (avg)
248.84 tok/s
TTFT (avg)
0.2 s
VRAM (avg)
17.64 GB
MTP acceptance rate
—
TTFB
0.39 GB
Power draw
72.4 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3-4B-Thinking-2507
Quantization
FP16
Framework
vLLM
Version
v0.24
Context length
8192 tokens
Batch size
8
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 M1 Max
Measured VRAM (avg)
17.64 GB
OS
macOS 15.0
Driver
Apple M1 Max
CUDA
N/A
Power draw
72.4 W

Sources & evidence

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

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

    v0.24 · macOS 15.0 · CUDA N/A · 8192 ctx

    48.27 tok/s

    Decode

    248.84 tok/s

    Prefill

    0.2 s

    TTFT

    17.64 GB

    VRAM

    —

    MTP

    72.4 W

    Power draw

    L2 batch seed 2026-10-16: vLLM on hw_id=101 running model_id=57 at FP16