Maker · GeForce RTX

NVIDIA

Best measured result

Qwen3.8-27B UD-Q4_K_XL

2.2× faster

Typical app 42.9

Our engine 93.3 t/s

measured on the GeForce RTX 3090

Every NVIDIA card we measured, before and after our engine. Same card, same model, same script.

Measured

Before and after on the same card, the same model and the same script.

NVIDIA GeForce RTX

3090
2.2×
- 24 GB memory
- 936 GB/s bandwidth
- Ampere sm_86
Qwen3.8-27B UD-Q4_K_XLbest model

Typical app 42.9

Our engine 93.3 t/s
NVIDIA GeForce RTX

5060 Ti
1.8×
- 16 GB memory
- 448 GB/s bandwidth
- Blackwell sm_120
Qwen3.8-27B UD-Q3_K_XLbest model

Typical app 28.1

Our engine + card profile 51.7 t/s
NVIDIA GeForce RTX

3060
1.6×
- 12 GB memory
- 360 GB/s bandwidth
- Ampere sm_86
Underdog Saluki 27B IQ2-mixbest model

Typical app 24.0

Our engine 38.2 t/s
NVIDIA GeForce RTX

4060 Ti
1.9×
- 16 GB memory
- 288 GB/s bandwidth
- Ada sm_89
Qwen3.8-27B UD-Q3_K_XLbest model

Typical app 18.6

Our engine 35.4 t/s

In the measurement queue

estimate

Not measured yet. The spec and, where possible, an estimate from our formula: not a measurement.

Take on another card

Two measured cards face to face, the same model.

Open the comparison