DeepSeek-V4-Flash · UD-IQ2_XXS
This page aggregates 1 real-world runs of DeepSeek-V4-Flash (UD-IQ2_XXS) on 4x NVIDIA RTX 3090 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.
35.68 tok/s
Decode
Decode speed
420.73 tok/s
Prefill
Prefill speed
0.11 s
TTFT
Time to first token
87.16 GB
VRAM
VRAM usage
1
Measured runs
1
Independent sources
GitHub
Source platforms
2026-09-27
Last verified
Performance
- llama.cpp · UD-IQ2_XXS (current)Decode 35.68 · Prefill 420.73 ·
Core figures
- Decode (avg)
- 35.68 tok/s
- Prefill (avg)
- 420.73 tok/s
- TTFT (avg)
- 0.11 s
- VRAM (avg)
- 87.16 GB
- MTP acceptance rate
- —
- TTFB
- — GB
- Power draw
- — W
Configuration
Member-level fields are taken from the most recent run
- Model
- DeepSeek-V4-Flash
- Quantization
- UD-IQ2_XXS
- Framework
- llama.cpp
- Version
- club3090-moecache-e0a2089
- 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
- 4x NVIDIA RTX 3090
- Nominal VRAM
- 96 GB
- Measured VRAM (avg)
- 87.16 GB
- OS
- Ubuntu 24.04.4 LTS
- Driver
- 595.84
- CUDA
- 13.2
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
club3090-moecache-e0a2089 · Ubuntu 24.04.4 LTS · CUDA 13.2 · 204800 ctx
35.68 tok/s
Decode
420.73 tok/s
Prefill
0.11 s
TTFT
87.16 GB
VRAM
—
MTP
— W
Power draw
4×3090 +3975WX/251GiB RAM，DeepSeek-V4-Flash-0731 IQ2_XXS、专家CPU卸载。narrative均值35.68t/s、短prompt TTFT111ms；9980tok prefill420.73t/s；同次bench峰值89250MiB≈87.16GiB。各卡250W功率限制；591.93W只是末尾瞬时总和，已置空。