Arao AI
Back to search results

Qwen3-4B · Q2_K

Start a discussion
Apple M1 Maxllama.cpp v0.46L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3-4B (Q2_K) on Apple M1 Max with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

145.16 tok/s

Decode

Decode speed

623.88 tok/s

Prefill

Prefill speed

2.44 s

TTFT

Time to first token

6.72 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 · Q2_K (current)Decode 145.16 · Prefill 623.88 ·
  2. Ollama · IQ2_M Decode 126.39 · Prefill 412.92 ·
  3. Ollama · Q8_0 Decode 64.73 · Prefill 325.03 ·

Core figures

Decode (avg)
145.16 tok/s
Prefill (avg)
623.88 tok/s
TTFT (avg)
2.44 s
VRAM (avg)
6.72 GB
MTP acceptance rate
—
TTFB
1.16 GB
Power draw
43.1 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3-4B
Quantization
Q2_K
Framework
llama.cpp
Version
v0.46
Context length
32768 tokens
Batch size
256
GPU layers
20
Flash Attention
Off

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)
6.72 GB
OS
macOS 15.0
Driver
Apple M1 Max
CUDA
N/A
Power draw
43.1 W

Sources & evidence

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

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

    v0.46 · macOS 15.0 · CUDA N/A · 32768 ctx

    145.16 tok/s

    Decode

    623.88 tok/s

    Prefill

    2.44 s

    TTFT

    6.72 GB

    VRAM

    —

    MTP

    43.1 W

    Power draw

    L2 batch seed 2026-10-16: llama.cpp on hw_id=101 running model_id=54 at Q2_K