Arao AI
Back to search results

Qwen3.5-0.8B · IQ2_M

Start a discussion
NVIDIA RTX 509032GBllama.cpp v0.30L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3.5-0.8B (IQ2_M) on NVIDIA RTX 5090 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

284.64 tok/s

Decode

Decode speed

1,115.95 tok/s

Prefill

Prefill speed

1.03 s

TTFT

Time to first token

2.97 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

2 days ago

Last verified

Performance

  1. llama.cpp · IQ2_M (current)Decode 284.64 · Prefill 1115.95 ·
  2. llama.cpp · Q4_K_M Decode 351.6 · Prefill 1360.11 ·
  3. vLLM · Q6_K Decode 341.56 · Prefill 1623.94 ·
  4. Ollama · Q3_K_L Decode 320.91 · Prefill 1586.97 ·
  5. Ollama · Q8_0 Decode 384.44 · Prefill 1342.7 ·
  6. Strata · FP16 Decode 337.54 · Prefill 1202.28 ·
  7. Strata · Q6_K Decode 317.49 · Prefill 1303.51 ·

Core figures

Decode (avg)
284.64 tok/s
Prefill (avg)
1,115.95 tok/s
TTFT (avg)
1.03 s
VRAM (avg)
2.97 GB
MTP acceptance rate
—
TTFB
2.28 GB
Power draw
362.4 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3.5-0.8B
Quantization
IQ2_M
Framework
llama.cpp
Version
v0.30
Context length
2048 tokens
Batch size
512
GPU layers
99
Flash Attention
Off

Hardware

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

GPU
NVIDIA RTX 5090
Nominal VRAM
32 GB
Measured VRAM (avg)
2.97 GB
OS
Ubuntu 22.04
Driver
570.86
CUDA
12.8
Power draw
362.4 W

Sources & evidence

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

  1. L2 Cross-framework verifiedOtherOriginal link Verified on 2026-10-16

    v0.30 · Ubuntu 22.04 · CUDA 12.8 · 2048 ctx

    284.64 tok/s

    Decode

    1,115.95 tok/s

    Prefill

    1.03 s

    TTFT

    2.97 GB

    VRAM

    —

    MTP

    362.4 W

    Power draw

    L2 batch seed 2026-10-16: llama.cpp on hw_id=13 running model_id=51 at IQ2_M