Arao AI
Back to search results

Qwen3-8B · Q6_K

Start a discussion
Apple M5 Ultra512GBllama.cpp v0.50L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3-8B (Q6_K) on Apple M5 Ultra with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

73.25 tok/s

Decode

Decode speed

373.92 tok/s

Prefill

Prefill speed

0.21 s

TTFT

Time to first token

18.9 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 · Q4_K_M Decode 102.09 · Prefill 450.54 ·
  2. llama.cpp · Q6_K (current)Decode 73.25 · Prefill 373.92 ·
  3. Ollama · IQ3_XS Decode 108 · Prefill 446.71 ·
  4. Ollama · Q4_K_S Decode 102.76 · Prefill 333.29 ·

Core figures

Decode (avg)
73.25 tok/s
Prefill (avg)
373.92 tok/s
TTFT (avg)
0.21 s
VRAM (avg)
18.9 GB
MTP acceptance rate
—
TTFB
0.59 GB
Power draw
119.7 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3-8B
Quantization
Q6_K
Framework
llama.cpp
Version
v0.50
Context length
2048 tokens
Batch size
512
GPU layers
33
Flash Attention
On

Hardware

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

GPU
Apple M5 Ultra
Nominal VRAM
512 GB
Measured VRAM (avg)
18.9 GB
OS
macOS 26.1
Power draw
119.7 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

    v0.50 · macOS 26.1 · 2048 ctx

    73.25 tok/s

    Decode

    373.92 tok/s

    Prefill

    0.21 s

    TTFT

    18.9 GB

    VRAM

    —

    MTP

    119.7 W

    Power draw

    L2 batch seed 2026-10-19: llama.cpp on hw_id=29 running model_id=53 at Q6_K