Pre-trained models for inference from Pi to iPhone

What it does

ailia-models is the official companion zoo for the commercial ailia inference SDK. It ships 403 pre-trained models—from YOLO and SAM to Whisper and Stable Diffusion—each with reference code and metadata showing which upstream project it came from and what SDK version it needs. Think of it as a translation layer that turns research checkpoints into something you can drop into a mobile app or embedded pipeline without touching an exporter.

The interesting bit

Most model zoos assume a Linux box with an NVIDIA GPU. This one targets a much wider matrix: Windows, macOS, iOS, Android, Jetson, and Raspberry Pi, using Vulkan and Metal for GPU acceleration instead of chaining you to CUDA. That breadth is the product; the models are just the inventory.

Key highlights

• 403 verified models as of March 2026 spanning detection, segmentation, LLMs, audio, 3D pose, and anomaly detection.
• Native GPU acceleration via Vulkan and Metal, with language bindings for Python, C++, C#, Kotlin, Rust, and Flutter.
• Each entry tracks its upstream source (mostly PyTorch), export provenance, and minimum ailia SDK version.
• Recently added models include SAM 2, Depth Anything V3, GPT-SoVITS, and Qwen2-VL.

Caveats

• Several newer int4-quantized models—notably whisper-int4, qwen2vl-int4, and gpt-sovits-v2pro-int4—currently require ONNX Runtime and do not run natively on the ailia SDK yet.

Verdict

Use this if you are already committed to the ailia SDK and need proven model integrations for edge or mobile deployment. If you want a framework-agnostic zoo or pure PyTorch playgrounds, look elsewhere.