On-device LLM Inference on Knowledge, Code, and Math prompts on Samsung Galaxy S25
57.8Average Accuracy (%)MobileMoE-L
Evaluation results
| Method | Links | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| MobileMoE-LMem (GB)=2.75, Backend=ExecuTorch+XNNPACK, Quantization (Weights/Activations)=INT4 weights, INT8 dynamic activations2026.05 | 57.8 | 1.53 | 4.09 | 8.99 | 25.85 | 55.12 | 162.45 | 53.4 | 43 | 36.6 | 22.9 | 17.3 | 8.9 | |
| MobileMoE-MMem (GB)=1.48, Backend=ExecuTorch+XNNPACK, Quantization (Weights/Activations)=INT4 weights, INT8 dynamic activations2026.05 | 52.5 | 0.84 | 1.79 | 4.32 | 16.68 | 35.02 | 88.25 | 83.6 | 77.1 | 64.6 | 34.5 | 26.2 | 15.4 | |
| MobileLLM-ProMem (GB)=0.55, Backend=ExecuTorch+XNNPACK, Quantization (Weights/Activations)=INT4 weights, INT8 dynamic activations2026.05 | 45.5 | 0.78 | 1.73 | 4.26 | 14.21 | 35.62 | 105.21 | 61.3 | 56.5 | 45.8 | 27.6 | 18.6 | 10.6 | |
| MobileMoE-SMem (GB)=0.68, Backend=ExecuTorch+XNNPACK, Quantization (Weights/Activations)=INT4 weights, INT8 dynamic activations2026.05 | 44 | 0.44 | 0.9 | 2.01 | 7.82 | 16.36 | 50.33 | 138.1 | 130 | 112 | 61.7 | 48.9 | 25.8 |