Arao AI
Back to search results

Llama-3.2-3B-Instruct · Q4_K_M

Start a discussion
NVIDIA RTX Spark128GBllama.cpp v0.27L2 Cross-framework verified

This page aggregates 1 real-world runs of Llama-3.2-3B-Instruct (Q4_K_M) on NVIDIA RTX Spark with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

89.55 tok/s

Decode

Decode speed

296 tok/s

Prefill

Prefill speed

1.73 s

TTFT

Time to first token

7.27 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_S Decode 108.82 · Prefill 395.74 ·
  2. llama.cpp · Q4_K_M (current)Decode 89.55 · Prefill 296 ·
  3. Ollama · IQ2_XS Decode 115.72 · Prefill 413.02 ·
  4. Ollama · Q4_K_M Decode 76.59 · Prefill 359.8 ·

Core figures

Decode (avg)
89.55 tok/s
Prefill (avg)
296 tok/s
TTFT (avg)
1.73 s
VRAM (avg)
7.27 GB
MTP acceptance rate
—
TTFB
0.72 GB
Power draw
84.6 W

Configuration

Member-level fields are taken from the most recent run

Model
Llama-3.2-3B-Instruct
Quantization
Q4_K_M
Framework
llama.cpp
Version
v0.27
Context length
16384 tokens
Batch size
512
GPU layers
20
Flash Attention
On

Hardware

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

GPU
NVIDIA RTX Spark
Nominal VRAM
128 GB
Measured VRAM (avg)
7.27 GB
OS
Windows 11 24H2
Driver
NVIDIA 590.10
CUDA
13.0
Power draw
84.6 W
Reference price
$2,399 · updated 2026-10-19

Prices are not live; check the store page

Sources & evidence

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

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

    v0.27 · Windows 11 24H2 · CUDA 13.0 · 16384 ctx

    89.55 tok/s

    Decode

    296 tok/s

    Prefill

    1.73 s

    TTFT

    7.27 GB

    VRAM

    —

    MTP

    84.6 W

    Power draw

    L2 batch seed 2026-10-19: llama.cpp on hw_id=110 running model_id=70 at Q4_K_M