qaihm-bot commited on
Commit
7a9a567
·
verified ·
1 Parent(s): b5dee5f
Files changed (3) hide show
  1. LICENSE +1 -0
  2. README.md +134 -0
  3. release_assets.json +94 -0
LICENSE ADDED
@@ -0,0 +1 @@
 
 
1
+ The license of the original trained model can be found at https://www.apache.org/licenses/LICENSE-2.0.
README.md ADDED
@@ -0,0 +1,134 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: pytorch
3
+ license: other
4
+ tags:
5
+ - llm
6
+ - vlm
7
+ - generative_ai
8
+ - android
9
+ pipeline_tag: text-generation
10
+
11
+ ---
12
+
13
+ ![](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/web-assets/model_demo.png)
14
+
15
+ # Qwen3-VL-4B-Instruct: Optimized for Qualcomm Devices
16
+
17
+ Qwen3-VL is a vision-language model from Alibaba Cloud capable of understanding both text and images for multimodal reasoning tasks such as visual question answering and image captioning.
18
+
19
+ This is based on the implementation of Qwen3-VL-4B-Instruct found [here](https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct).
20
+ 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/v0.57.0/src/qai_hub_models/models/qwen3_vl_4b_instruct) library to export with custom configurations. More details on model performance across various devices, can be found [here](#performance-summary).
21
+
22
+ 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.
23
+
24
+ ## Deploying Qwen3-VL-4B-Instruct on-device
25
+
26
+ Follow the [GenieX quickstart](https://geniex.aihub.qualcomm.com/en/get-started/quickstart) to install GenieX and deploy the model on a target device.
27
+
28
+ See the [LLM-on-Genie](https://github.com/qualcomm/ai-hub-apps/tree/main/tutorials/llm_on_genie) tutorial to run with the Genie runtime. Note: Genie support will be deprecated soon.
29
+
30
+ ## Getting Started
31
+ There are two ways to deploy this model on your device:
32
+
33
+ ### Option 1: Download Pre-Exported Models
34
+
35
+ Below are pre-exported model assets ready for deployment.
36
+
37
+ | Runtime | Precision | Chipset | SDK Versions | Download |
38
+ |---|---|---|---|---|
39
+ | GENIE | w4a16 | Snapdragon® 8 Elite Gen 5 Mobile | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-genie-w4a16-qualcomm_snapdragon_8_elite_gen5.zip)
40
+ | GENIE | w4a16 | Snapdragon® 8 Elite Mobile | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-genie-w4a16-qualcomm_snapdragon_8_elite.zip)
41
+ | GENIE | w4a16 | Snapdragon® X2 Elite | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-genie-w4a16-qualcomm_snapdragon_x2_elite.zip)
42
+ | GENIE | w4a16 | Snapdragon® X Elite | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-genie-w4a16-qualcomm_snapdragon_x_elite.zip)
43
+ | GENIE | w4a16 | Qualcomm® SA8775P | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-genie-w4a16-qualcomm_sa8775p.zip)
44
+ | GENIE | w4a16 | Qualcomm® QCS9075 | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-genie-w4a16-qualcomm_qcs9075.zip)
45
+ | GENIEX_LLAMACPP | q4_0 | Universal | | [Download](https://huggingface.co/unsloth/Qwen3-VL-4B-Instruct-GGUF/resolve/main/Qwen3-VL-4B-Instruct-Q4_0.gguf)
46
+ | GENIEX_QAIRT | w4a16 | Snapdragon® 8 Elite Gen 5 Mobile | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-geniex_qairt-w4a16-qualcomm_snapdragon_8_elite_gen5.zip)
47
+ | GENIEX_QAIRT | w4a16 | Snapdragon® 8 Elite Mobile | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-geniex_qairt-w4a16-qualcomm_snapdragon_8_elite.zip)
48
+ | GENIEX_QAIRT | w4a16 | Snapdragon® X2 Elite | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-geniex_qairt-w4a16-qualcomm_snapdragon_x2_elite.zip)
49
+ | GENIEX_QAIRT | w4a16 | Snapdragon® X Elite | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-geniex_qairt-w4a16-qualcomm_snapdragon_x_elite.zip)
50
+ | GENIEX_QAIRT | w4a16 | Qualcomm® QCS9075 | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-geniex_qairt-w4a16-qualcomm_qcs9075.zip)
51
+
52
+ For more device-specific assets and performance metrics, visit **[Qwen3-VL-4B-Instruct on Qualcomm® AI Hub](https://aihub.qualcomm.com/models/qwen3_vl_4b_instruct)**.
53
+
54
+
55
+ ### Option 2: Export with Custom Configurations
56
+
57
+ Use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/v0.57.0/src/qai_hub_models/models/qwen3_vl_4b_instruct) Python library to compile and export the model with your own:
58
+ - Custom weights (e.g., fine-tuned checkpoints)
59
+ - Custom input shapes
60
+ - Target device and runtime configurations
61
+
62
+ This option is ideal if you need to customize the model beyond the default configuration provided here.
63
+
64
+ See our repository for [Qwen3-VL-4B-Instruct on GitHub](https://github.com/qualcomm/ai-hub-models/blob/v0.57.0/src/qai_hub_models/models/qwen3_vl_4b_instruct) for usage instructions.
65
+
66
+ ## Model Details
67
+
68
+ **Model Type:** Model_use_case.text_generation
69
+
70
+ **Model Stats:**
71
+ - Model architecture: Transformer with ViT Vision Encoder, Grouped Query Attention (GQA), and SwiGLU activation.
72
+ - Supported languages: 100+ languages and dialects
73
+ - TTFT: Time To First Token is the time it takes to generate the first response token. This is expressed as a range because it varies based on the length of the prompt.
74
+ - Response Rate: Rate of response generation after the first response token.
75
+
76
+ ## Performance Summary
77
+ | Model | Runtime | Precision | Chipset | Context Length | Response Rate (tokens per second) | Time To First Token (range, seconds)
78
+ |---|---|---|---|---|---|---
79
+ | Qwen3-VL-4B-Instruct | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Gen 5 Mobile | 512 | 21.937586 | 1.04066225 - 4.162649
80
+ | Qwen3-VL-4B-Instruct | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Gen 5 Mobile | 512 | 21.063662 | 1.1498887500000001 - 4.5995550000000005
81
+ | Qwen3-VL-4B-Instruct | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Gen 5 Mobile | 512 | 12.51382 | 0.17715475 - 0.708619
82
+ | Qwen3-VL-4B-Instruct | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Gen 5 Mobile | 4096 | 7.594744 | 3.0647713125 - 98.072682
83
+ | Qwen3-VL-4B-Instruct | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Gen 5 Mobile | 4096 | 7.347485 | 3.0688408437500003 - 98.20290700000001
84
+ | Qwen3-VL-4B-Instruct | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Gen 5 Mobile | 4096 | 12.036943 | 0.33230665625 - 10.633813
85
+ | Qwen3-VL-4B-Instruct | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Mobile | 512 | 21.038925 | 1.15949925 - 4.637997
86
+ | Qwen3-VL-4B-Instruct | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Mobile | 512 | 20.872003 | 1.22482375 - 4.899295
87
+ | Qwen3-VL-4B-Instruct | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Mobile | 512 | 18.041792 | 0.24882474999999998 - 0.9952989999999999
88
+ | Qwen3-VL-4B-Instruct | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Mobile | 4096 | 5.266429 | 3.2057110625 - 102.582754
89
+ | Qwen3-VL-4B-Instruct | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Mobile | 4096 | 5.404354 | 3.23921925 - 103.655016
90
+ | Qwen3-VL-4B-Instruct | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Mobile | 4096 | 11.884297 | 0.49755121874999997 - 15.921638999999999
91
+ | Qwen3-VL-4B-Instruct | GENIEX_LLAMACPP | q4_0 | Snapdragon® X Elite | 512 | 21.892482 | 0.4739375 - 1.89575
92
+ | Qwen3-VL-4B-Instruct | GENIEX_LLAMACPP | q4_0 | Snapdragon® X Elite | 512 | 22.078185 | 0.51991475 - 2.079659
93
+ | Qwen3-VL-4B-Instruct | GENIEX_LLAMACPP | q4_0 | Snapdragon® X Elite | 512 | 11.538787 | 0.37511574999999997 - 1.5004629999999999
94
+ | Qwen3-VL-4B-Instruct | GENIEX_LLAMACPP | q4_0 | Snapdragon® X Elite | 4096 | 7.683006 | 1.2810328125000001 - 40.993050000000004
95
+ | Qwen3-VL-4B-Instruct | GENIEX_LLAMACPP | q4_0 | Snapdragon® X Elite | 4096 | 8.547017 | 1.3233025 - 42.34568
96
+ | Qwen3-VL-4B-Instruct | GENIEX_LLAMACPP | q4_0 | Snapdragon® X Elite | 4096 | 7.576911 | 0.5775706562499999 - 18.482260999999998
97
+ | Qwen3-VL-4B-Instruct | GENIEX_QAIRT | w4a16 | Snapdragon® 8 Elite Gen 5 Mobile | 4096 | 29.162899 | 0.26334399999999997 - 4.2135039999999995
98
+ | Qwen3-VL-4B-Instruct | GENIEX_QAIRT | w4a16 | Snapdragon® 8 Elite Mobile | 4096 | 27.22378 | 0.317968 - 10.174976
99
+ | Qwen3-VL-4B-Instruct | GENIEX_QAIRT | w4a16 | Snapdragon® X2 Elite | 4096 | 39.217323 | 0.24981866666666666 - 7.994197333333333
100
+ | Qwen3-VL-4B-Instruct | GENIEX_QAIRT | w4a16 | Snapdragon® X Elite | 4096 | 20.886665 | 0.5332319999999999 - 17.063423999999998
101
+ | Qwen3-VL-4B-Instruct | GENIEX_QAIRT | w4a16 | Qualcomm® QCS9075 | 4096 | 18.401298 | 0.48866666666666664 - 15.637333333333332
102
+
103
+ ## License
104
+ * The license for the original implementation of Qwen3-VL-4B-Instruct can be found
105
+ [here](https://www.apache.org/licenses/LICENSE-2.0).
106
+
107
+ ## References
108
+ * [Qwen3 Technical Report](https://arxiv.org/abs/2505.09388)
109
+ * [Source Model Implementation](https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct)
110
+
111
+ ## Community
112
+ * Join [our AI Hub Slack community](https://aihub.qualcomm.com/community/slack) to collaborate, post questions and learn more about on-device AI.
113
+ * For questions or feedback please [reach out to us](mailto:ai-hub-support@qti.qualcomm.com).
114
+
115
+ ## Usage and Limitations
116
+
117
+ This model may not be used for or in connection with any of the following applications:
118
+
119
+ - Accessing essential private and public services and benefits;
120
+ - Administration of justice and democratic processes;
121
+ - Assessing or recognizing the emotional state of a person;
122
+ - Biometric and biometrics-based systems, including categorization of persons based on sensitive characteristics;
123
+ - Education and vocational training;
124
+ - Employment and workers management;
125
+ - Exploitation of the vulnerabilities of persons resulting in harmful behavior;
126
+ - General purpose social scoring;
127
+ - Law enforcement;
128
+ - Management and operation of critical infrastructure;
129
+ - Migration, asylum and border control management;
130
+ - Predictive policing;
131
+ - Real-time remote biometric identification in public spaces;
132
+ - Recommender systems of social media platforms;
133
+ - Scraping of facial images (from the internet or otherwise); and/or
134
+ - Subliminal manipulation
release_assets.json ADDED
@@ -0,0 +1,94 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": "0.57.0",
3
+ "precisions": {
4
+ "q4_0": {
5
+ "universal_assets": {
6
+ "geniex_llamacpp": {
7
+ "download_url": "https://huggingface.co/unsloth/Qwen3-VL-4B-Instruct-GGUF/resolve/main/Qwen3-VL-4B-Instruct-Q4_0.gguf"
8
+ }
9
+ }
10
+ },
11
+ "w4a16": {
12
+ "chipset_assets": {
13
+ "qualcomm-snapdragon-8-elite": {
14
+ "genie": {
15
+ "tool_versions": {
16
+ "qairt": "2.45.0.260326154327"
17
+ },
18
+ "download_url": "https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-genie-w4a16-qualcomm_snapdragon_8_elite.zip"
19
+ },
20
+ "geniex_qairt": {
21
+ "tool_versions": {
22
+ "qairt": "2.45.0.260326154327"
23
+ },
24
+ "download_url": "https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-geniex_qairt-w4a16-qualcomm_snapdragon_8_elite.zip"
25
+ }
26
+ },
27
+ "qualcomm-qcs9075": {
28
+ "genie": {
29
+ "tool_versions": {
30
+ "qairt": "2.45.0.260326154327"
31
+ },
32
+ "download_url": "https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-genie-w4a16-qualcomm_qcs9075.zip"
33
+ },
34
+ "geniex_qairt": {
35
+ "tool_versions": {
36
+ "qairt": "2.45.0.260326154327"
37
+ },
38
+ "download_url": "https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-geniex_qairt-w4a16-qualcomm_qcs9075.zip"
39
+ }
40
+ },
41
+ "qualcomm-sa8775p": {
42
+ "genie": {
43
+ "tool_versions": {
44
+ "qairt": "2.45.0.260326154327"
45
+ },
46
+ "download_url": "https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-genie-w4a16-qualcomm_sa8775p.zip"
47
+ }
48
+ },
49
+ "qualcomm-snapdragon-8-elite-gen5": {
50
+ "genie": {
51
+ "tool_versions": {
52
+ "qairt": "2.45.0.260326154327"
53
+ },
54
+ "download_url": "https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-genie-w4a16-qualcomm_snapdragon_8_elite_gen5.zip"
55
+ },
56
+ "geniex_qairt": {
57
+ "tool_versions": {
58
+ "qairt": "2.45.0.260326154327"
59
+ },
60
+ "download_url": "https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-geniex_qairt-w4a16-qualcomm_snapdragon_8_elite_gen5.zip"
61
+ }
62
+ },
63
+ "qualcomm-snapdragon-x-elite": {
64
+ "genie": {
65
+ "tool_versions": {
66
+ "qairt": "2.45.0.260326154327"
67
+ },
68
+ "download_url": "https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-genie-w4a16-qualcomm_snapdragon_x_elite.zip"
69
+ },
70
+ "geniex_qairt": {
71
+ "tool_versions": {
72
+ "qairt": "2.45.0.260326154327"
73
+ },
74
+ "download_url": "https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-geniex_qairt-w4a16-qualcomm_snapdragon_x_elite.zip"
75
+ }
76
+ },
77
+ "qualcomm-snapdragon-x2-elite": {
78
+ "genie": {
79
+ "tool_versions": {
80
+ "qairt": "2.45.0.260326154327"
81
+ },
82
+ "download_url": "https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-genie-w4a16-qualcomm_snapdragon_x2_elite.zip"
83
+ },
84
+ "geniex_qairt": {
85
+ "tool_versions": {
86
+ "qairt": "2.45.0.260326154327"
87
+ },
88
+ "download_url": "https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_vl_4b_instruct/releases/v0.57.0/qwen3_vl_4b_instruct-geniex_qairt-w4a16-qualcomm_snapdragon_x2_elite.zip"
89
+ }
90
+ }
91
+ }
92
+ }
93
+ }
94
+ }