Arao AI
Back to search results

Qwen3.5-4B · Q3_K_L

Start a discussion
Apple M2 Ultrallama.cpp v0.51L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3.5-4B (Q3_K_L) on Apple M2 Ultra with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

147.11 tok/s

Decode

Decode speed

678.92 tok/s

Prefill

Prefill speed

1.62 s

TTFT

Time to first token

8.15 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

16 days ago

Last verified

Performance

  1. llama.cpp · IQ3_XS Decode 156.36 · Prefill 856.45 ·
  2. llama.cpp · Q3_K_L (current)Decode 147.11 · Prefill 678.92 ·
  3. Ollama · IQ2_M Decode 176 · Prefill 698.29 ·
  4. vLLM · Q3_K_L Decode 164.19 · Prefill 535.64 ·
  5. vLLM · Q3_K_S Decode 169.83 · Prefill 644.07 ·
  6. Ollama · FP16 Decode 60.62 · Prefill 204.54 ·

Core figures

Decode (avg)
147.11 tok/s
Prefill (avg)
678.92 tok/s
TTFT (avg)
1.62 s
VRAM (avg)
8.15 GB
MTP acceptance rate
—
TTFB
1.18 GB
Power draw
123.3 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3.5-4B
Quantization
Q3_K_L
Framework
llama.cpp
Version
v0.51
Context length
4096 tokens
Batch size
512
GPU layers
99
Flash Attention
Off

Hardware

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

GPU
Apple M2 Ultra
Measured VRAM (avg)
8.15 GB
OS
macOS 26.1
Power draw
123.3 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

    v0.51 · macOS 26.1 · 4096 ctx

    147.11 tok/s

    Decode

    678.92 tok/s

    Prefill

    1.62 s

    TTFT

    8.15 GB

    VRAM

    —

    MTP

    123.3 W

    Power draw

    L2 batch seed 2026-10-19: llama.cpp on hw_id=103 running model_id=48 at Q3_K_L