DeepSeek-V4-Flash-Vision-Exp · UD-Q8_K_XL
This page aggregates 1 real-world runs of DeepSeek-V4-Flash-Vision-Exp (UD-Q8_K_XL) on 2x NVIDIA RTX 5090 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.
32.08 tok/s
Decode
Decode speed
671.45 tok/s
Prefill
Prefill speed
0.12 s
TTFT
Time to first token
60.59 GB
VRAM
VRAM usage
1
Measured runs
1
Independent sources
GitHub
Source platforms
2026-09-27
Last verified
Performance
- llama.cpp · UD-Q8_K_XL (current)Decode 32.08 · Prefill 671.45 ·
Core figures
- Decode (avg)
- 32.08 tok/s
- Prefill (avg)
- 671.45 tok/s
- TTFT (avg)
- 0.12 s
- VRAM (avg)
- 60.59 GB
- MTP acceptance rate
- —
- TTFB
- — GB
- Power draw
- — W
Configuration
Member-level fields are taken from the most recent run
- Model
- DeepSeek-V4-Flash-Vision-Exp
- Quantization
- UD-Q8_K_XL
- Framework
- llama.cpp
- Version
- moecachev1.6-rc0
- Context length
- 204800 tokens
- GPU layers
- 99
Hardware
Nominal and measured figures are shown side by side; whether it runs is the reader's call
- GPU
- 2x NVIDIA RTX 5090
- Nominal VRAM
- 64 GB
- Measured VRAM (avg)
- 60.59 GB
- OS
- Ubuntu 24.04.4 LTS
- Driver
- 610.57.04
- 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.
Sources & evidence
1 measured records in total, each traceable to its original source
moecachev1.6-rc0 · Ubuntu 24.04.4 LTS · CUDA 13.3 · 204800 ctx
32.08 tok/s
Decode
671.45 tok/s
Prefill
0.12 s
TTFT
60.59 GB
VRAM
—
MTP
— W
Power draw
2×5090 +3975WX/251GiB RAM，MoE专家CPU卸载。narrative5次decode均值32.08t/s，短prompt TTFT119ms；cache-busted10K prefill均值671.45t/s，bench峰值62047MiB≈60.59GiB。专家缓存未启用，不能据此宣称cache加速。UD-Q8_K_XL为混合精度，日志MXFP4 MoE不等于整模型纯8bit。