Bring Your Own Model
Convert your own model with the Torq compiler and run it on the SL2600 NPU.
Convert your own model with the Torq compiler and run it on the SL2600 NPU.
A multimodal embedding model from Google that maps text, images, video and audio into one 768-dimensional space for cross-modal search, optimized for the Torq NPU.
A three-stage on-device face recognition pipeline: face detection, 68-point facial keypoints, and 256-D face embeddings.
Turn spoken or typed commands into hardware actions with FunctionGemma on the SL2610.
A lightweight 270M parameter model from Google, built on Gemma 3 270M and trained for function calling, intended as a foundation for specialized fine-tuned models.
A compact 270M parameter multimodal language model from Google optimized for text classification and data extraction tasks with rapid fine-tuning capability.
A compact 270M parameter language model from Google, quantized to INT4, for text classification and data extraction tasks with rapid fine-tuning capability.
Set up your board and run your first edge AI examples.
Recognize and label images with a classifier on the SL2600 NPU.
A 450M parameter vision-language model from Liquid AI that answers natural-language questions about an input image.
A 230M parameter hybrid short-convolution and attention language model from Liquid AI that can act as a skill-selection layer, decomposing instructions into tool calls.
Translate speech into other languages with Moonshine and Gemma 3 on the SL2610.
A quantized MobileNet V2 model with a 1.0 width multiplier optimized for image classification on ImageNet at 224x224 resolution.
A streaming automatic speech recognition model with a sliding-window encoder for low time-to-first-token, optimized for 16kHz mono audio.
A high-efficiency automatic speech recognition model using variable-length architecture for real-time speech recognition, optimized for 16kHz mono audio.
Turn object detections into ambient music with YOLOv8 and MelodyRNN on the SL2610.
Find and locate objects in images with YOLO on the SL2600 NPU.
A 0.5B parameter instruction-tuned language model from the Qwen team, quantized to INT4 with GPTQ for the Torq NPU.
A 0.6B parameter language model from the Qwen team, quantized to INT4 with GPTQ for the Torq NPU.
Transcribe speech from a microphone with Moonshine on the SL2610 NPU.
Live word-by-word transcription with Moonshine Streaming on the SL2610 NPU.
An overview of the Torq Edge AI platform for running neural networks on the SL2610 NPU.
Describe what the camera sees with the LFM2-VL vision-language model on the SL2610 NPU.
YOLO26 NMS-free object detection models (nano) optimized for 320x320 resolution, with the decode head compiled onto the NPU.
YOLO26 NMS-free object detection models (small) optimized for 320x320 resolution, with the decode head compiled onto the NPU.
A YOLOv8n model for object detection optimized for 320x320 resolution, offering a balance between speed and accuracy.
A YOLOv8n pose estimation model optimized for 320x320 resolution, detecting people and their body keypoints.
A YOLOv8s pose estimation model optimized for 320x320 resolution, detecting people and their body keypoints.
A YOLOv8s model for object detection optimized for 320x320 resolution, offering a balance between speed and accuracy.