TFLite(LiteRT) - Analytical AI

The resources in this section support both the Online and Offline inference paths for TFLite-analytical AI workloads. This section provides the necessary TFLite models and pre-compiled Deep Learning Archives (DLAs) required to execute models via the TFLite Interpreter or directly through the Neuron Runtime.

Note

To prepare your own models for these paths (such as quantization or compilation), refer to AI Development Resources to obtain the required NeuroPilot SDKs and host tools.

Vision models in the TFLite Analytical Model collection perform three core tasks on image data: image classification, object detection, and recognition. Together, these tasks help identify what appears in an image, where it is, and who or what a specific entity is. The Analytical Model section provides a curated set of commonly used TensorFlow Lite models and their converted Deep Learning Archives (DLAs) for deployment on MediaTek Genio platforms.

Each model package includes the following artifacts:

• A baseline TensorFlow Lite (.tflite) model.

• One or more version-specific DLA binaries:

  • mdla2 for MDLA 2.x,

  • mdla3 for MDLA 3.x, and

  • mdla5 for MDLA 5.x.

You must run each DLA variant only on platforms that support the corresponding MDLA version.

MDLA Version Mapping

Use the following table to look up the MDLA version and the hardware accelerator supported by each Genio platform and select the matching DLA model:

MDLA Version and Supported Hardware Accelerator

Platform	OS	MDLA Version	TFLite - Analytical AI (Online)	TFLite - Analytical AI (Offline)
Genio 520/720	Android	5.3	CPU + GPU + NPU	NPU
	Yocto	5.3	CPU + GPU + NPU	NPU
Genio 510/700	Android	3	CPU + GPU + NPU	NPU
	Yocto	3	CPU + GPU + NPU	NPU
	Ubuntu	3	CPU + GPU + NPU	NPU
Genio 1200	Android	2	CPU + GPU + NPU	NPU
	Yocto	2	CPU + GPU + NPU	NPU
	Ubuntu	2	CPU + GPU + NPU	NPU
Genio 350	Android	X	CPU + GPU + NPU	X
	Yocto	X	CPU + GPU	X
	Ubuntu	X	CPU + GPU	X

Vision Tasks and Model Categories

The TFLite Analytical Model collection focuses on three primary visual AI tasks: image classification, object detection, and recognition. Each model family page documents the conversion workflow and benchmark performance for representative models on MediaTek Genio platforms.

Image Classification

Image classification is a fundamental computer vision task that predicts the category or class of an input image. Unlike object detection, classification models do not provide object locations. They output only the most likely label for the entire image.

For TFLite Analytical image classification models, the most common backbones are the following families on Genio platforms:

• DenseNet
• EfficientNet
• MobileNetV2
• MobileNetV3
• ResNet
• SqueezeNet
• VGG

Object Detection

Object detection is a key task in computer vision that identifies and localizes objects within an image. Unlike classification, it outputs both the class and the bounding box of each detected object.

For TFLite Analytical object detection models, the most common backbones are the following families on Genio platforms:

• YOLOv5
• YOLOv8

Recognition

Recognition models identify or verify specific entities from input data, such as faces, persons, gestures, or text. Unlike classification, which predicts a general category, recognition focuses on determining who or what a target object is. These models often compare the target against a known database of identities or reference templates.

For TFLite Analytical recognition models, the most common backbones are the following families on Genio platforms:

• VGGFace

Community Contributed Models

In addition to the primary task categories, the collection includes Community Contributed Models that are provided by the community. These models are shared without the original source training models, and model provenance or training recipes might not be available.

• Community Contributed Model

The available analytical and community contributed models for each task and Genio platform are listed in the tables below.

Supported Models on Genio Products

The following tables list the supported analytical models per Genio platform for each task category.

Note

The performance statistics shown in these tables were measured using offline inference with performance mode enabled across different Genio products, models, and data types.

Models for Detection

					Genio 520	Genio 720	Genio 510	Genio 700	Genio 1200	MT8893
Task	Model Name	Source model type	Data Type	Input Size	Inference Time (ms)	Inference Time (ms)	Inference Time (ms)	Inference Time (ms)	Inference Time (ms)	Inference Time (ms)	Detail
Object Detection	YOLOv5s	.pt	Quant8	640x640	5.35	5.39	17.47	10.04	19.05	3.42	Link
Object Detection	YOLOv5s	.pt	Float32	640x640	16.37	16.23	46.41	32.04	36.66	11.4	Link
Object Detection	YOLOv8s	.pt	Quant8	640x640	7.85	7.63	25.51	17.01	28.04	5.64	Link
Object Detection	YOLOv8s	.pt	Float32	640x640	24.22	31.07	70.95	50.04	55.84	16.34	Link

Models for Classification

					Genio 520	Genio 720	Genio 510	Genio 700	Genio 1200	MT8893
Task	Model Name	Source model type	Data Type	Input Size	Inference Time (ms)	Inference Time (ms)	Inference Time (ms)	Inference Time (ms)	Inference Time (ms)	Inference Time (ms)	Detail
Classification	DenseNet	.pt	Quant8	224x224	3.68	4.08	7.03	5.03	6.04	2.4	Link
Classification	DenseNet	.pt	Float32	224x224	7.51	8.41	16.51	11.04	12.04	4.74	Link
Classification	EfficientNet	.pt	Quant8	224x224	1.49	1.9	4.05	3	3.05	1.16	Link
Classification	EfficientNet	.pt	Float32	224x224	3.03	3.73	9.03	6.04	6.05	2.21	Link
Classification	MobileNetV2	.pt	Quant8	224x224	1.04	1.19	1.37	1.04	1.04	0.78	Link
Classification	MobileNetV2	.pt	Float32	224x224	1.89	2.17	3.57	2.04	2.58	1.29	Link
Classification	MobileNetV3	.pt	Quant8	224x224	0.73	1.04	1.04	0.04	N/A	0.64	Link
Classification	MobileNetV3	.pt	Float32	224x224	1.19	1.56	2.72	1.05	2.05	0.97	Link
Classification	ResNet	.pt	Quant8	224x224	1.5	1.65	2.79	2.03	2.05	1.08	Link
Classification	ResNet	.pt	Float32	224x224	3.82	5.19	9.21	6.04	8.04	2.56	Link
Classification	SqueezeNet	.pt	Quant8	224x224	1.19	1.77	1.52	1.04	1.05	0.86	Link
Classification	SqueezeNet	.pt	Float32	224x224	2.19	3.19	5.01	3.04	3.05	1.64	Link
Classification	VGG	.pt	Quant8	224x224	10.91	12.74	24.85	17.04	24.04	6.47	Link
Classification	VGG	.pt	Float32	224x224	33.42	43.34	80.3	56.04	49.05	19.87	Link

Models for Recognition

					Genio 520	Genio 720	Genio 510	Genio 700	Genio 1200	MT8893
Task	Model Name	Source model type	Data Type	Input Size	Inference Time (ms)	Inference Time (ms)	Inference Time (ms)	Inference Time (ms)	Inference Time (ms)	Inference Time (ms)	Detail
Recognition	VGGFace	.pt	Quant8	224x224	11.23	13.34	25.04	17.04	24.04	6.58	Link
Recognition	VGGFace	.pt	Float32	224x224	34.05	44.35	81.7	56.04	49.05	20	Link

Performance Notes and Limitations

Note

The measurements were obtained using Neuron SDK(neuronrt), and each model’s detail page (linked from the tables) provides additional information about the test setup and configuration.

Performance can vary depending on:

• The specific Genio platform and hardware configuration.

• The version of the board image and evaluation kit (EVK).

• The selected backend and model variant.

To obtain the most accurate performance numbers for your use case, you must run the application directly on the target platform.

Important

For online inference on Yocto, some models may not run on certain backends due to custom operators generated by the MediaTek converter. These custom operators (for example, MTK_EXT ops) are not recognized or supported by the standard TensorFlow Lite interpreter, which can lead to incompatibility issues during inference. In such cases, the corresponding entries in the tables are marked as N/A to indicate unavailable data.

The exact cause of a failure or unsupported configuration may vary per model. For more details, refer to the model‑specific documentation.