Hugging Face's logo

qualcomm
/
MiniLM-v2

Text Generation
PyTorch
foundation
real_time
android
Model card Files
xet
 Community
MiniLM-v2
File size: 12,730 Bytes
7218b03
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
288efa9
7218b03
 
 
 
 
 
 
 
 
 
 
 
288efa9
 
 
 
 
 
7218b03
 
 
 
 
 
288efa9
7218b03
 
 
 
 
 
288efa9
7218b03
 
 
 
 
 
 
 
 
 
 
 
 
 
 
288efa9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7218b03
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
---
library_name: pytorch
license: other
tags:
- foundation
- real_time
- android
pipeline_tag: text-generation

---

![](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/minilm_v2/web-assets/model_demo.png)

# MiniLM-v2: Optimized for Qualcomm Devices

All-MiniLM-L6-v2 maps sentences to a 384-dimensional dense vector space. Trained on 1B+ sentence pairs, it excels at semantic search, clustering, and sentence similarity tasks while being small enough to run on mobile devices.

This is based on the implementation of MiniLM-v2 found [here](https://github.com/UKPLab/sentence-transformers).
This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/v0.58.0/src/qai_hub_models/models/minilm_v2) library to export with custom configurations. More details on model performance across various devices, can be found [here](#performance-summary).

Qualcomm AI Hub Models uses [Qualcomm AI Hub Workbench](https://workbench.aihub.qualcomm.com) to compile, profile, and evaluate this model. [Sign up](https://myaccount.qualcomm.com/signup) to run these models on a hosted Qualcomm® device.

## Getting Started
There are two ways to deploy this model on your device:

### Option 1: Download Pre-Exported Models

Below are pre-exported model assets ready for deployment.

| Runtime | Precision | Chipset | SDK Versions | Download |
|---|---|---|---|---|
| ONNX | float | Universal | QAIRT 2.45, ONNX Runtime 1.25.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/minilm_v2/releases/v0.58.0/minilm_v2-onnx-float.zip)
| ONNX | w8a8 | Universal | QAIRT 2.45, ONNX Runtime 1.25.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/minilm_v2/releases/v0.58.0/minilm_v2-onnx-w8a8.zip)
| QNN_DLC | float | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/minilm_v2/releases/v0.58.0/minilm_v2-qnn_dlc-float.zip)
| QNN_DLC | w8a8 | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/minilm_v2/releases/v0.58.0/minilm_v2-qnn_dlc-w8a8.zip)
| TFLITE | float | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/minilm_v2/releases/v0.58.0/minilm_v2-tflite-float.zip)
| TFLITE | w8a8 | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/minilm_v2/releases/v0.58.0/minilm_v2-tflite-w8a8.zip)

For more device-specific assets and performance metrics, visit **[MiniLM-v2 on Qualcomm® AI Hub](https://aihub.qualcomm.com/models/minilm_v2)**.


### Option 2: Export with Custom Configurations

Use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/v0.58.0/src/qai_hub_models/models/minilm_v2) Python library to compile and export the model with your own:
- Custom weights (e.g., fine-tuned checkpoints)
- Custom input shapes
- Target device and runtime configurations

This option is ideal if you need to customize the model beyond the default configuration provided here.

See our repository for [MiniLM-v2 on GitHub](https://github.com/qualcomm/ai-hub-models/blob/v0.58.0/src/qai_hub_models/models/minilm_v2) for usage instructions.

## Model Details

**Model Type:** Model_use_case.text_generation

**Model Stats:**
- Model checkpoint: sentence-transformers/all-MiniLM-L6-v2
- Input resolution: 128 tokens
- Number of parameters: 22.7M
- Model size (float): 86.7 MB
- Embedding dimension: 384

## Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit
|---|---|---|---|---|---|---
| MiniLM-v2 | ONNX | float | Snapdragon® X2 Elite | 0.781 ms | 1 - 1 MB | NPU
| MiniLM-v2 | ONNX | float | Snapdragon® X Elite | 1.751 ms | 45 - 45 MB | NPU
| MiniLM-v2 | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 1.227 ms | 0 - 104 MB | NPU
| MiniLM-v2 | ONNX | float | Snapdragon® 8 Gen 1 Mobile | 2.547 ms | 0 - 108 MB | NPU
| MiniLM-v2 | ONNX | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 1.71 ms | 0 - 3 MB | NPU
| MiniLM-v2 | ONNX | float | Qualcomm® QCS8450 | 2.547 ms | 0 - 108 MB | NPU
| MiniLM-v2 | ONNX | float | Qualcomm® Dragonwing™ IQ-9075 | 2.074 ms | 0 - 3 MB | NPU
| MiniLM-v2 | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.7 ms | 0 - 62 MB | NPU
| MiniLM-v2 | ONNX | float | Snapdragon® 8 Elite Mobile | 0.864 ms | 0 - 60 MB | NPU
| MiniLM-v2 | ONNX | float | Qualcomm® Dragonwing™ Q-8750 | 0.864 ms | 0 - 60 MB | NPU
| MiniLM-v2 | ONNX | float | Qualcomm® Dragonwing™ IQ-X7181 | 1.751 ms | 45 - 45 MB | NPU
| MiniLM-v2 | ONNX | w8a8 | Snapdragon® X2 Elite | 0.673 ms | 1 - 1 MB | NPU
| MiniLM-v2 | ONNX | w8a8 | Snapdragon® X Elite | 1.783 ms | 23 - 23 MB | NPU
| MiniLM-v2 | ONNX | w8a8 | Snapdragon® 8 Gen 3 Mobile | 1.202 ms | 0 - 85 MB | NPU
| MiniLM-v2 | ONNX | w8a8 | Snapdragon® 8 Gen 1 Mobile | 1.973 ms | 0 - 87 MB | NPU
| MiniLM-v2 | ONNX | w8a8 | Qualcomm® Dragonwing™ QCS6490 | 5.522 ms | 0 - 3 MB | NPU
| MiniLM-v2 | ONNX | w8a8 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 1.704 ms | 0 - 2 MB | NPU
| MiniLM-v2 | ONNX | w8a8 | Qualcomm® QCS8450 | 1.973 ms | 0 - 87 MB | NPU
| MiniLM-v2 | ONNX | w8a8 | Snapdragon® 8 Elite Mobile | 0.845 ms | 0 - 57 MB | NPU
| MiniLM-v2 | ONNX | w8a8 | Snapdragon® 7 Gen 4 Mobile | 1.939 ms | 0 - 55 MB | NPU
| MiniLM-v2 | ONNX | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.666 ms | 0 - 62 MB | NPU
| MiniLM-v2 | ONNX | w8a8 | Qualcomm® Dragonwing™ IQ-9075 | 1.8 ms | 0 - 3 MB | NPU
| MiniLM-v2 | ONNX | w8a8 | Qualcomm® Dragonwing™ Q-6690 | 6.703 ms | 0 - 61 MB | NPU
| MiniLM-v2 | ONNX | w8a8 | Qualcomm® Dragonwing™ Q-7790 | 1.939 ms | 0 - 55 MB | NPU
| MiniLM-v2 | ONNX | w8a8 | Qualcomm® Dragonwing™ Q-8750 | 0.845 ms | 0 - 57 MB | NPU
| MiniLM-v2 | ONNX | w8a8 | Qualcomm® Dragonwing™ IQ-X7181 | 1.783 ms | 23 - 23 MB | NPU
| MiniLM-v2 | QNN_DLC | float | Snapdragon® X2 Elite | 0.672 ms | 1 - 1 MB | NPU
| MiniLM-v2 | QNN_DLC | float | Snapdragon® X Elite | 1.3 ms | 1 - 1 MB | NPU
| MiniLM-v2 | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 0.797 ms | 0 - 100 MB | NPU
| MiniLM-v2 | QNN_DLC | float | Snapdragon® 8 Gen 1 Mobile | 1.888 ms | 0 - 100 MB | NPU
| MiniLM-v2 | QNN_DLC | float | Qualcomm® QCS8275 | 3.554 ms | 0 - 58 MB | NPU
| MiniLM-v2 | QNN_DLC | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 1.111 ms | 0 - 20 MB | NPU
| MiniLM-v2 | QNN_DLC | float | Qualcomm® SA8775P | 1.513 ms | 0 - 60 MB | NPU
| MiniLM-v2 | QNN_DLC | float | Qualcomm® SA8650P | 1.513 ms | 0 - 60 MB | NPU
| MiniLM-v2 | QNN_DLC | float | Qualcomm® SA8255P | 1.513 ms | 0 - 60 MB | NPU
| MiniLM-v2 | QNN_DLC | float | Qualcomm® QCS8450 | 1.888 ms | 0 - 100 MB | NPU
| MiniLM-v2 | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-9075 | 1.389 ms | 0 - 2 MB | NPU
| MiniLM-v2 | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.459 ms | 0 - 61 MB | NPU
| MiniLM-v2 | QNN_DLC | float | Qualcomm® SA7255P | 3.554 ms | 0 - 58 MB | NPU
| MiniLM-v2 | QNN_DLC | float | Snapdragon® 8 Elite Mobile | 0.563 ms | 0 - 59 MB | NPU
| MiniLM-v2 | QNN_DLC | float | Qualcomm® SA8295P | 2.293 ms | 0 - 56 MB | NPU
| MiniLM-v2 | QNN_DLC | float | Qualcomm® Dragonwing™ Q-8750 | 0.563 ms | 0 - 59 MB | NPU
| MiniLM-v2 | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-X7181 | 1.3 ms | 1 - 1 MB | NPU
| MiniLM-v2 | QNN_DLC | w8a8 | Snapdragon® X2 Elite | 0.8 ms | 0 - 0 MB | NPU
| MiniLM-v2 | QNN_DLC | w8a8 | Snapdragon® X Elite | 1.903 ms | 0 - 0 MB | NPU
| MiniLM-v2 | QNN_DLC | w8a8 | Snapdragon® 8 Gen 3 Mobile | 1.211 ms | 3 - 72 MB | NPU
| MiniLM-v2 | QNN_DLC | w8a8 | Snapdragon® 8 Gen 1 Mobile | 1.944 ms | 0 - 74 MB | NPU
| MiniLM-v2 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ QCS6490 | 4.196 ms | 2 - 4 MB | NPU
| MiniLM-v2 | QNN_DLC | w8a8 | Qualcomm® QCS8275 | 3.465 ms | 0 - 48 MB | NPU
| MiniLM-v2 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 1.727 ms | 0 - 15 MB | NPU
| MiniLM-v2 | QNN_DLC | w8a8 | Qualcomm® SA8775P | 1.951 ms | 0 - 51 MB | NPU
| MiniLM-v2 | QNN_DLC | w8a8 | Qualcomm® SA8650P | 1.951 ms | 0 - 51 MB | NPU
| MiniLM-v2 | QNN_DLC | w8a8 | Qualcomm® SA8255P | 1.951 ms | 0 - 51 MB | NPU
| MiniLM-v2 | QNN_DLC | w8a8 | Qualcomm® QCS8450 | 1.944 ms | 0 - 74 MB | NPU
| MiniLM-v2 | QNN_DLC | w8a8 | Snapdragon® 8 Elite Mobile | 0.832 ms | 0 - 48 MB | NPU
| MiniLM-v2 | QNN_DLC | w8a8 | Snapdragon® 7 Gen 4 Mobile | 1.701 ms | 0 - 47 MB | NPU
| MiniLM-v2 | QNN_DLC | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.677 ms | 0 - 50 MB | NPU
| MiniLM-v2 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ IQ-9075 | 1.821 ms | 0 - 2 MB | NPU
| MiniLM-v2 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ Q-6690 | 6.131 ms | 0 - 50 MB | NPU
| MiniLM-v2 | QNN_DLC | w8a8 | Qualcomm® SA7255P | 3.465 ms | 0 - 48 MB | NPU
| MiniLM-v2 | QNN_DLC | w8a8 | Qualcomm® SA8295P | 2.357 ms | 0 - 46 MB | NPU
| MiniLM-v2 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ Q-7790 | 1.701 ms | 0 - 47 MB | NPU
| MiniLM-v2 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ Q-8750 | 0.832 ms | 0 - 48 MB | NPU
| MiniLM-v2 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ IQ-X7181 | 1.903 ms | 0 - 0 MB | NPU
| MiniLM-v2 | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 0.819 ms | 0 - 106 MB | NPU
| MiniLM-v2 | TFLITE | float | Snapdragon® 8 Gen 1 Mobile | 1.966 ms | 0 - 102 MB | NPU
| MiniLM-v2 | TFLITE | float | Qualcomm® QCS8275 | 3.676 ms | 0 - 63 MB | NPU
| MiniLM-v2 | TFLITE | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 1.117 ms | 0 - 2 MB | NPU
| MiniLM-v2 | TFLITE | float | Qualcomm® SA8775P | 1.565 ms | 0 - 64 MB | NPU
| MiniLM-v2 | TFLITE | float | Qualcomm® SA8650P | 1.565 ms | 0 - 64 MB | NPU
| MiniLM-v2 | TFLITE | float | Qualcomm® SA8255P | 1.565 ms | 0 - 64 MB | NPU
| MiniLM-v2 | TFLITE | float | Qualcomm® QCS8450 | 1.966 ms | 0 - 102 MB | NPU
| MiniLM-v2 | TFLITE | float | Qualcomm® Dragonwing™ IQ-9075 | 1.408 ms | 0 - 45 MB | NPU
| MiniLM-v2 | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.487 ms | 0 - 58 MB | NPU
| MiniLM-v2 | TFLITE | float | Qualcomm® SA7255P | 3.676 ms | 0 - 63 MB | NPU
| MiniLM-v2 | TFLITE | float | Snapdragon® 8 Elite Mobile | 0.597 ms | 0 - 60 MB | NPU
| MiniLM-v2 | TFLITE | float | Qualcomm® SA8295P | 2.361 ms | 0 - 56 MB | NPU
| MiniLM-v2 | TFLITE | float | Qualcomm® Dragonwing™ Q-8750 | 0.597 ms | 0 - 60 MB | NPU
| MiniLM-v2 | TFLITE | w8a8 | Snapdragon® 8 Gen 3 Mobile | 1.668 ms | 0 - 80 MB | NPU
| MiniLM-v2 | TFLITE | w8a8 | Snapdragon® 8 Gen 1 Mobile | 2.624 ms | 0 - 80 MB | NPU
| MiniLM-v2 | TFLITE | w8a8 | Qualcomm® Dragonwing™ QCS6490 | 16.748 ms | 0 - 31 MB | NPU
| MiniLM-v2 | TFLITE | w8a8 | Qualcomm® QCS8275 | 4.624 ms | 0 - 54 MB | NPU
| MiniLM-v2 | TFLITE | w8a8 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 2.367 ms | 0 - 4 MB | NPU
| MiniLM-v2 | TFLITE | w8a8 | Qualcomm® SA8775P | 2.653 ms | 0 - 59 MB | NPU
| MiniLM-v2 | TFLITE | w8a8 | Qualcomm® SA8650P | 2.653 ms | 0 - 59 MB | NPU
| MiniLM-v2 | TFLITE | w8a8 | Qualcomm® SA8255P | 2.653 ms | 0 - 59 MB | NPU
| MiniLM-v2 | TFLITE | w8a8 | Qualcomm® QCS8450 | 2.624 ms | 0 - 80 MB | NPU
| MiniLM-v2 | TFLITE | w8a8 | Snapdragon® 8 Elite Mobile | 1.136 ms | 0 - 61 MB | NPU
| MiniLM-v2 | TFLITE | w8a8 | Snapdragon® 7 Gen 4 Mobile | 5.071 ms | 0 - 35 MB | NPU
| MiniLM-v2 | TFLITE | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.947 ms | 0 - 53 MB | NPU
| MiniLM-v2 | TFLITE | w8a8 | Qualcomm® Dragonwing™ IQ-9075 | 2.51 ms | 0 - 24 MB | NPU
| MiniLM-v2 | TFLITE | w8a8 | Qualcomm® Dragonwing™ Q-6690 | 11.97 ms | 0 - 37 MB | NPU
| MiniLM-v2 | TFLITE | w8a8 | Qualcomm® SA7255P | 4.624 ms | 0 - 54 MB | NPU
| MiniLM-v2 | TFLITE | w8a8 | Qualcomm® SA8295P | 3.209 ms | 0 - 48 MB | NPU
| MiniLM-v2 | TFLITE | w8a8 | Qualcomm® Dragonwing™ Q-7790 | 5.071 ms | 0 - 35 MB | NPU
| MiniLM-v2 | TFLITE | w8a8 | Qualcomm® Dragonwing™ Q-8750 | 1.136 ms | 0 - 61 MB | NPU

## License
* The license for the original implementation of MiniLM-v2 can be found
  [here](https://github.com/UKPLab/sentence-transformers/blob/master/LICENSE).

## References
* [Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks](https://arxiv.org/abs/1908.10084)
* [Source Model Implementation](https://github.com/UKPLab/sentence-transformers)

## Community
* Join [our AI Hub Slack community](https://aihub.qualcomm.com/community/slack) to collaborate, post questions and learn more about on-device AI.
* For questions or feedback please [reach out to us](mailto:ai-hub-support@qti.qualcomm.com).