Arao AI
Back to search results

Kimi-K3 · IQ4_XS

Start a discussion
NVIDIA RTX 409024GBllama.cpp v0.44L2 Cross-framework verified

This page aggregates 1 real-world runs of Kimi-K3 (IQ4_XS) on NVIDIA RTX 4090 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

123.9 tok/s

Decode

Decode speed

634.46 tok/s

Prefill

Prefill speed

0.58 s

TTFT

Time to first token

1.05 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

Today

Last verified

Performance

  1. llama.cpp · IQ4_XS (current)Decode 123.9 · Prefill 634.46 ·
  2. Ollama · Q3_K_L Decode 113.94 · Prefill 395.79 ·
  3. Ollama · Q8_0 Decode 70.56 · Prefill 253.27 ·
  4. SGLang · Q5_K_M Decode — · Prefill — ·
  5. SGLang · Q4_K_M Decode — · Prefill — ·
  6. vLLM · Q5_K_M Decode — · Prefill — ·
  7. vLLM · Q4_K_M Decode — · Prefill — ·

Core figures

Decode (avg)
123.9 tok/s
Prefill (avg)
634.46 tok/s
TTFT (avg)
0.58 s
VRAM (avg)
1.05 GB
MTP acceptance rate
—
TTFB
0.99 GB
Power draw
355.7 W

Configuration

Member-level fields are taken from the most recent run

Model
Kimi-K3
Quantization
IQ4_XS
Framework
llama.cpp
Version
v0.44
Context length
4096 tokens
Batch size
256
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 4090
Nominal VRAM
24 GB
Measured VRAM (avg)
1.05 GB
OS
Ubuntu 24.04
Driver
570.86
CUDA
13.0
Power draw
355.7 W

Sources & evidence

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

  1. L2 Cross-framework verifiedOtherOriginal link Verified on 2026-10-18

    v0.44 · Ubuntu 24.04 · CUDA 13.0 · 4096 ctx

    123.9 tok/s

    Decode

    634.46 tok/s

    Prefill

    0.58 s

    TTFT

    1.05 GB

    VRAM

    —

    MTP

    355.7 W

    Power draw

    L2 batch seed 2026-10-18: llama.cpp on hw_id=1 running model_id=73 at IQ4_XS