Samtool Supported Models May 2026

✓ Model: efficientnet_b0.onnx ✓ Operators (24 total): Conv, Relu, Add, GlobalAvgPool, Reshape ⚠ Unsupported on nvidia_gpu: HardSwish (fallback to CPU) ✓ Memory: 342 MB (fit) Result: PARTIAL_SUPPORT (1 unsupported op) If your model uses a custom operator, you can register it via Samtool’s plugin interface. To illustrate the practical impact of Samtool’s optimization, here are inference latency numbers (ms per batch=1) on a Qualcomm Snapdragon 8 Gen 2 (Hexagon DSP) for key supported models:

In the rapidly evolving landscape of artificial intelligence and machine learning, efficient hardware exploitation is no longer a luxury—it is a necessity. For developers, data scientists, and system administrators working with inference and deployment, the toolchain that bridges the gap between AI models and physical hardware is critical. One such powerful, though often under-documented, tool in this ecosystem is Samtool . samtool supported models

| Model | TFLite (ms) | Samtool optimized (ms) | Speedup | |-------|-------------|------------------------|---------| | MobileNetV3 | 4.2 | 2.1 | 2.0x | | ResNet-50 | 23.5 | 12.8 | 1.84x | | BERT Base | 78.3 | 41.2 | 1.90x | | YOLOv8 Nano | 14.7 | 8.3 | 1.77x | | Whisper Tiny | 112.5 | 67.4 | 1.67x | ✓ Model: efficientnet_b0

The true value of Samtool lies not just in the model list, but in the hardware portability it delivers. A single command can take a PyTorch ResNet-50 and deploy it to an NVIDIA GPU, a Qualcomm NPU in a smartphone, or an ARM Cortex-M microcontroller—all without rewriting a single line of application code. One such powerful, though often under-documented, tool in

| Input Format | Version Support | Notes | |--------------|----------------|-------| | | opset 9 through 18 | Most complete support. Recommended for new models. | | TensorFlow Lite | v1.15 to v2.14 | Specialized for edge models. | | PyTorch (TorchScript) | torch 1.9 to 2.2 | Via torch.jit.trace or script . | | TensorFlow SavedModel | v2.x | Limited; prefer ONNX conversion. | | Keras H5 | v2.6+ | Experimental. | | PaddlePaddle | via X2Paddle conversion | Community-supported. | | Core ML | macOS target only | For Apple Neural Engine (ANE). |