Arao AI
Back to search results

DeepSeek-R1-Distill-Qwen-14B · Q4_K_M

Start a discussion
NVIDIA H2097GBllama.cpp v0.45L2 Cross-framework verified

This page aggregates 1 real-world runs of DeepSeek-R1-Distill-Qwen-14B (Q4_K_M) on NVIDIA H20 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

42.42 tok/s

Decode

Decode speed

221.34 tok/s

Prefill

Prefill speed

0.14 s

TTFT

Time to first token

26.21 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 · IQ2_XS Decode 67.98 · Prefill 225.25 ·
  2. llama.cpp · Q4_K_M (current)Decode 42.42 · Prefill 221.34 ·
  3. vLLM · Q2_K Decode 62.74 · Prefill 220.68 ·
  4. vLLM · Q6_K Decode 33.53 · Prefill 137.11 ·

Core figures

Decode (avg)
42.42 tok/s
Prefill (avg)
221.34 tok/s
TTFT (avg)
0.14 s
VRAM (avg)
26.21 GB
MTP acceptance rate
—
TTFB
0.34 GB
Power draw
303.6 W

Configuration

Member-level fields are taken from the most recent run

Model
DeepSeek-R1-Distill-Qwen-14B
Quantization
Q4_K_M
Framework
llama.cpp
Version
v0.45
Context length
16384 tokens
Batch size
512
GPU layers
20
Flash Attention
Off

Hardware

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

GPU
NVIDIA H20
Nominal VRAM
97 GB
Measured VRAM (avg)
26.21 GB
OS
Ubuntu 24.04
Driver
560.94
CUDA
12.8
Power draw
303.6 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

    v0.45 · Ubuntu 24.04 · CUDA 12.8 · 16384 ctx

    42.42 tok/s

    Decode

    221.34 tok/s

    Prefill

    0.14 s

    TTFT

    26.21 GB

    VRAM

    —

    MTP

    303.6 W

    Power draw

    L2 batch seed 2026-10-19: llama.cpp on hw_id=56 running model_id=71 at Q4_K_M