Arao AI
Back to search results

Qwen3-4B-Thinking-2507 · FP16

Start a discussion
Apple M1vLLM v0.53L2 Cross-framework verified

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

30.09 tok/s

Decode

Decode speed

100.41 tok/s

Prefill

Prefill speed

0.41 s

TTFT

Time to first token

15.2 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 · IQ2_XS Decode 75.23 · Prefill 381.59 ·
  2. llama.cpp · Q4_K_M Decode 59.36 · Prefill 212.58 ·
  3. vLLM · FP16 (current)Decode 30.09 · Prefill 100.41 ·
  4. vLLM · IQ2_XS Decode 79.87 · Prefill 436.11 ·

Core figures

Decode (avg)
30.09 tok/s
Prefill (avg)
100.41 tok/s
TTFT (avg)
0.41 s
VRAM (avg)
15.2 GB
MTP acceptance rate
—
TTFB
0.24 GB
Power draw
18.8 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3-4B-Thinking-2507
Quantization
FP16
Framework
vLLM
Version
v0.53
Context length
8192 tokens
Batch size
16
GPU layers
999
Flash Attention
Off

Hardware

Nominal and measured figures are shown side by side; whether it runs is the reader's call

GPU
Apple M1
Measured VRAM (avg)
15.2 GB
OS
macOS 15.0
Driver
Apple M1
CUDA
N/A
Power draw
18.8 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

    v0.53 · macOS 15.0 · CUDA N/A · 8192 ctx

    30.09 tok/s

    Decode

    100.41 tok/s

    Prefill

    0.41 s

    TTFT

    15.2 GB

    VRAM

    —

    MTP

    18.8 W

    Power draw

    L2 batch seed 2026-10-16: vLLM on hw_id=99 running model_id=57 at FP16