Darml

Model in. Firmware out.

The model-to-firmware compiler for edge AI. Upload TFLite, ONNX, or scikit-learn — get a flashable binary for AVR, STM32, ESP32, Raspberry Pi, or Jetson. Single CLI. No vendor lock-in.

$pip install darml
$darml build model.tflite --target esp32-s3
$darml flash darml-build.zip --port /dev/ttyUSB0
$screen /dev/ttyUSB0 115200
[Darml] Ready.
[Darml] pred=2 conf=0.94 latency=12µs
[Darml] pred=2 conf=0.93 latency=11µs
# Model running on hardware. Total wall time: ~6 minutes (first build).

For

Embedded engineers who already have a trained model and need it running on a real device. ML researchers shipping a demo to a chip. Teams replacing custom toolchain glue with one binary.

Not for

Training new models — bring your own. Cloud-only inference — we ship to silicon. AutoML — we don't pick architectures, we deploy yours.

From your model to your device, in five steps.

No notebooks. No configuration files. No vendor SDKs. The CLI takes your model file and your target board, tells you whether it'll fit, and hands you something flashable.

01 / UPLOAD

Hand it your model

.tflite, .onnx, or .pkl — Darml auto-detects the format.

02 / PICK

Pick your board

11 supported targets across 5 hardware tiers. darml targets lists them all.

03 / CHECK

Get warned before you wait

Will it fit in flash? Will it fit in RAM? You find out before the toolchain spins up.

04 / BUILD

Get a binary

firmware.bin for MCUs, .tar.gz for Linux boards. Plus a manifest with target-specific flash instructions.

05 / FLASH

Run it

darml flash <artifact.zip> --port … — auto-detects the right tool: esptool, STM32_Programmer_CLI, avrdude, ssh+docker.

Real models, real targets, real numbers.

Every release runs through a 5-model × 11-target acceptance matrix (55 parse + size-check combinations) plus end-to-end firmware builds. These are the actual outputs from the latest passing run.

Model
Target
Format → output
Output size
Build time
micro_speech (1D-Conv kw spotter)
esp32-s3
tflite → firmware.bin
566 KB
273 s
MNIST CNN (Conv2D + Dense)
stm32f4
tflite → firmware.bin
345 KB
176 s
MLP (multi-layer perceptron)
esp32-s3
onnx → tflite → firmware.bin
522 KB
270 s
Random forest (iris)
avr-mega328
sklearn → C library
1.5 KB
~1 s
MobileNetV2 INT8
rpi5
tflite → tarball + Dockerfile
597 KB
~2 s

Plus 73 passing unit tests covering parser, size estimator, pipeline, build cache, license validation, and inference roundtrip (FP32 ONNX vs INT8 TFLite, argmax-stable on ≥ 80% of samples). Run the same matrix yourself: ./scripts/pre_publish.sh.

11 targets. 5 hardware tiers.

AVR for the smallest sensors, Cortex-M for industrial control, Xtensa for connected devices, Linux-class for vision and audio. One CLI across all of them — no vendor SDK switching.

2–8 KB RAM

AVR (Arduino-class)

emlearn · pure C · <2 KB at runtime
avr-mega328
avr-mega2560
320 KB – 1.5 MB RAM

STM32 (Cortex-M)

TFLite Micro · ARM GCC
stm32f4 · stm32h7 · stm32n6
520 KB – 8.5 MB RAM

ESP32 (Xtensa)

TFLite Micro · WiFi/BT
esp32 · esp32-s3 (PSRAM)
4–8 GB RAM

Raspberry Pi

TFLite · Python + Docker
rpi4 · rpi5
4–8 GB RAM

NVIDIA Jetson

TensorRT / TFLite · Docker
jetson-nano · jetson-orin
Don't see your board? If it runs PlatformIO, we can probably add it. Open an issue with the chip + dev-board name and we'll prioritize.

Pricing

Free to start. Self-host the same code your customers are running. Industrial buyers prefer self-hosted annual licenses; we honor that.

Core

€0 / forever

MIT — fully open source. Use it for anything, commercial or not.

  • All 11 hardware targets
  • CLI with serial output
  • 5 builds/day (per-machine)
  • Community support
pip install darml

Pro Cloud

€49 / month

We host the build farm. No PlatformIO, no toolchain setup, no maintenance.

  • Unlimited builds
  • INT8 auto-quantization
  • ONNX → TFLite auto-convert
  • Web dashboard + shared cache
  • HTTP / MQTT result output
  • Email support
Start free trial

Pro Self-Hosted

€499 / year / seat

Run it inside your network. Air-gapped factories, secure dev environments, on-prem CI.

  • Everything in Pro Cloud
  • Runs offline / air-gapped
  • Offline license key (HMAC)
  • 14-day free trial built in
  • Commercial license
  • Email support, 1-day SLA
Start free trial

Enterprise

Custom

Dedicated build farm, signed firmware, SOC 2, fleet-management add-ons.

  • Everything in Pro Self-Hosted
  • Signed firmware artifacts
  • Dedicated build runners
  • SSO / SAML
  • Volume seat pricing
  • Phone + email SLA
Talk to us