Smaller, more efficient edge AI for billions of devices.
The platform to deploy and benchmark AI models on real microcontrollers.
Smart camerasAudio devicesWearablesRobotics & drones
From model to microcontroller.
Bring a model – upload a TFLite file, or export from PyTorch with our open-source SDK. DeepGate compiles it, benchmarks it on real hardware, and gets it running on your device.
Deploy to a wide range of targets.
Select from 100s of microcontrollers and let the DeepGate compiler turn your model into a ready-to-flash binary – up to 2× faster and using 3× less RAM than TFLite Micro.
Measure on real hardware. Build smaller, faster models.
Profile your model on real microcontrollers with a per-layer performance breakdown. See exactly where the time and memory go, then iterate.
Build in PyTorch. Ship with DeepGate.
The DeepGate SDK extends PyTorch with everything needed to build efficient AI for microcontrollers. Train, quantize, export, and deploy – all from a single workflow.
- Build in PyTorch – use standard PyTorch with DeepGate layers for maximum efficiency
- Built-in preprocessing – image and audio preprocessing exported with your model
- Train & quantize – integrated QAT and PTQ, no separate quantization tools
- Export – directly export deployable DeepGate models
import dg
import torch.nn as nn
from dg.preprocess import AudioPipeline, PeakNormalize, Window
# Build with standard PyTorch
model = nn.Sequential(nn.Conv2d(1, 16, 3), nn.ReLU(), dg.Flatten())
# Preprocessing exported with the model
preprocess = AudioPipeline([PeakNormalize(), Window(480, 320)])
# Quantize
quantized = dg.post_training_quantize(model, calib_dataset)
# Export and upload to app.deepgate.ai
schema = dg.export(quantized, preprocess=preprocess)Built for efficiency.
DeepGate Compiler was built to deliver maximum AI efficiency, replacing the TensorFlow Lite Micro (TFLM) runtime used by most MCU vendors – bringing AI to constrained hardware.
- Zero manual tuning – op registration and memory allocation handled automatically by DeepGate
- Automatic memory planning – whole-graph buffer reuse for up to 3× less RAM
- Target-specific kernels – optimized per MCU for up to 2× faster inference
Need AI in your product?
We build it with you. Bring your use case and we’ll deliver a model that’s trained, optimised and validated for your hardware.
Already have a model?We can make it up to 45× faster using up to 11× less RAM.