0
25
50
75
100
accuracy (%)
2L · 25M
4L · 29M
8L · 52M
16L · 98M
20L · 121M
base and tuned scored with forced calls · DeepSeek V4 Flash through its cloud API
Every subnetwork, fine-tuned on DroidCall
1 of 2
60.5
20.5
56.5
34.5
62.5
42.5
68.0
49.5
69.0
52.0
70.0
Needle 3 base
fine-tuned on DroidCall
DeepSeek V4 Flash
Every subnetwork, fine-tuned on Mobile Actions
2 of 2
88.4
0.9
66.7
19.0
79.1
45.2
83.7
80.2
84.6
86.7
84.5
Needle 3 base
fine-tuned on Mobile Actions
DeepSeek V4 Flash