Model Farm Guide

Model Farm is the model marketplace for AidLux. It provides open-source models optimized for Qualcomm edge chipsets, runnable preprocessing, postprocessing, and inference examples, and performance figures measured on the stated hardware. Browsing models and performance data does not require an account; downloading model packages and example code does.

Model Farm supports Qualcomm QCS6490, QCS8550, QCS8625, Dragonwing IQ9, and Dragonwing IQ8 platforms. Downloaded models can run with AidLux through AidLite, AidGen, or Qualcomm QNN.

2.1 Selecting a Model

Use model type, precision, chipset, and keyword filters to locate a model. A model detail page provides the following evaluation data:

Field	Description
Device	Development board and chipset used for the measurement.
AI framework	Framework and version used for model conversion and inference.
Model precision	Such as FP16, INT8, or W4A16.
Inference latency	Measured inference time excluding preprocessing and postprocessing.
Accuracy loss	Cosine similarity comparison between the converted model output and the FP32 source model output.
Model size	Size of the converted model file.

Performance on another board with the same SoC is for reference only. Always download an entry whose chipset, precision, and QNN backend version exactly match the target environment.

2.2 Downloading with MMS

MMS (Model Management Service) lets developers query and download models from a board terminal. Register an account on the developer registration page, then sign in:

mms login

List all models or filter by keyword:

mms list
mms list yolo

The output identifies the model name, precision, chipset, and backend version. Use those exact values in the download command:

# -m: model name  -p: precision  -c: chipset  -b: QNN backend version  -d: download directory
mms get -m yolov6l -p int8 -c qcs8550 -b qnn2.23 -d /home/aidlux/yolov6l

For example, download an FP16 QNN 2.36 YOLOv8s package for IQ9/QCS9075:

mms list | grep YOLOv8s | grep FP16
mms get -m YOLOv8s -p fp16 -c iq9 -b qnn2.36 -d ./
unzip YOLOv8s_iq9_fp16.zip

Always use the actual local mms list output. Package filenames, internal directory names, and chipset aliases in examples can differ. After extraction, follow the package README.md rather than inferring compatibility from its filename.

2.3 Running Downloaded Vision Models

Vision-model packages downloaded through MMS generally have this structure:

{model_name}_{SoC}_{precision}/
|-- models/       # Converted model files
|-- code/
|   |-- python/   # AidLite Python examples
|   `-- cpp/      # AidLite C++ examples
`-- README.md

For example, the QCS8550 YOLO11s-obb package can use its AidLite Python example for NPU inference:

cd code
sudo python3 python/run_test.py \
  --target_model ../models/QCS8550/FP16/yolo11s-obb_qcs8550_fp16.qnn236.ctx.bin \
  --imgs python/boats.jpg \
  --invoke_nums 10

Model Farm includes models for object detection, classification, monocular depth estimation, segmentation, pose estimation, and oriented bounding-box detection. Example entries include YOLOv8s, ConvNeXt-Tiny, Depth-Anything-V2-Small, FastSAM-S, YOLO11l-Pose, and YOLO11s-obb.

2.4 Running Generative Models

AidGen loads generative models: use aidllm for language models and aidmlm for vision-language models. An MMS package normally contains one or more .aidem shards plus cache, embedding, or vision-model files required by its configuration.

Build the language-model example project from the package directory:

cp -r /usr/local/share/aidgen/examples/aidgen_qnn240/cpp/aidllm/ ./
cd aidllm
mkdir build && cd build
cmake ..
make

From the model directory, run the application with the package configuration or a configuration created for that model:

./aidllm/build/test_aidllm abort ./config.json

Available models vary by platform and QNN backend. The source tutorials include examples for MiniCPM5-1B, Qwen3-8B, Meta-Llama-3.1-8B-Instruct, Gemma-2-2B-it, Falcon3-7B-Instruct, DeepSeek-R1-Distill-Qwen-7B, Phi-3.5-mini-instruct, HY-MT1.5-1.8B, and Qwen2.5-VL-3B-Instruct. Shard count, context length, prompt template, and configuration fields vary by model; use the corresponding package README and configuration example.

For HTTP integration, use AidGenSE instead: query models with aidllm remote-list api, download one with aidllm pull api <model-address>, then start an OpenAI-compatible service with aidllm start api -m <model-name>.

2.5 Fine-Tuned and Preview Models

For a fine-tuned model, refer to Model Conversion Reference on the model detail page or in the package README. Use AIMO to convert the model to a Qualcomm platform format, then replace the model in the example project with the generated .amf file for testing.

Preview models marked Contact Us are not available for direct download from the website. Obtain these models through MMS on an APLUX development board and run them with AidLite SDK.