Arao AI
Back to search results

Qwen3.6-27B · IQ3_XS

Start a discussion
Apple M5 Ultra512GBllama.cpp v0.21L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3.6-27B (IQ3_XS) on Apple M5 Ultra with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

91.11 tok/s

Decode

Decode speed

454.84 tok/s

Prefill

Prefill speed

1.45 s

TTFT

Time to first token

37.25 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

17 days ago

Last verified

Performance

  1. llama.cpp · IQ3_XS (current)Decode 91.11 · Prefill 454.84 ·
  2. llama.cpp · Q4_K_S Decode 81.06 · Prefill 387.21 ·
  3. Ollama · IQ3_S Decode 85.88 · Prefill 388.53 ·
  4. Ollama · Q6_K Decode 53.29 · Prefill 218.93 ·

Core figures

Decode (avg)
91.11 tok/s
Prefill (avg)
454.84 tok/s
TTFT (avg)
1.45 s
VRAM (avg)
37.25 GB
MTP acceptance rate
—
TTFB
0.73 GB
Power draw
127.2 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3.6-27B
Quantization
IQ3_XS
Framework
llama.cpp
Version
v0.21
Context length
4096 tokens
Batch size
512
GPU layers
28
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)
37.25 GB
OS
macOS 26.1
Power draw
127.2 W

Get started

Original model resources

Base model resources may not match this quantization. Use the verified deployment weights above when available.

Sources & evidence

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

  1. L2 Cross-framework verifiedOtherOriginal link Verified on 2026-10-19

    v0.21 · macOS 26.1 · 4096 ctx

    91.11 tok/s

    Decode

    454.84 tok/s

    Prefill

    1.45 s

    TTFT

    37.25 GB

    VRAM

    —

    MTP

    127.2 W

    Power draw

    L2 batch seed 2026-10-19: llama.cpp on hw_id=29 running model_id=6 at IQ3_XS