Qwen3.6-27B · Q5_K_S
This page aggregates 1 real-world runs of Qwen3.6-27B (Q5_K_S) on NVIDIA RTX 3090 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.
46.57 tok/s
Decode
Decode speed
1,201.22 tok/s
Prefill
Prefill speed
0.1 s
TTFT
Time to first token
21.53 GB
VRAM
VRAM usage
1
Measured runs
1
Independent sources
GitHub
Source platforms
2026-09-27
Last verified
Performance
- vLLM · IQ1_S Decode 96.12 · Prefill 373.79 ·
- Ollama · IQ1_S Decode 76.57 · Prefill 351.59 ·
- llama.cpp · Q4_K_M Decode 57.02 · Prefill 1299.01 ·
- llama.cpp · UD-Q3_K_XL Decode 36.67 · Prefill 39.44 ·
- llama.cpp · Q5_K_S (current)Decode 46.57 · Prefill 1201.22 ·
Core figures
- Decode (avg)
- 46.57 tok/s
- Prefill (avg)
- 1,201.22 tok/s
- TTFT (avg)
- 0.1 s
- VRAM (avg)
- 21.53 GB
- MTP acceptance rate
- —
- TTFB
- — GB
- Power draw
- — W
Configuration
Member-level fields are taken from the most recent run
- Model
- Qwen3.6-27B
- Quantization
- Q5_K_S
- Framework
- llama.cpp
- Version
- beellama.cpp v0.3.0-e0663be2713c
- Context length
- 102400 tokens
Hardware
Nominal and measured figures are shown side by side; whether it runs is the reader's call
- GPU
- NVIDIA RTX 3090
- Nominal VRAM
- 24 GB
- Measured VRAM (avg)
- 21.53 GB
- OS
- Debian GNU/Linux 13 (trixie)
- Driver
- 610.43.02
- CUDA
- 13.3
Get started
Weights for this deployment
Original model resources
Base model resources may not match this quantization. Use the verified deployment weights above when available.
Deployment guide
https://github.com/Anbeeld/beellama.cppSources & evidence
1 measured records in total, each traceable to its original source
beellama.cpp v0.3.0-e0663be2713c · Debian GNU/Linux 13 (trixie) · CUDA 13.3 · 102400 ctx
46.57 tok/s
Decode
1,201.22 tok/s
Prefill
0.1 s
TTFT
21.53 GB
VRAM
—
MTP
— W
Power draw
RTX3090 + i7-7800X/48GiB RAM，beellama.cpp DFlash fork、Q5_K_S，ctx102400。narrative5次均值46.57t/s，短prompt TTFT98ms；13179tok prefill1201.22t/s。21.53GiB为初始idle读数，非峰值；soak因容器探测失败未完成，不能当长期稳定性证明。