darml 0.1.2

Model-to-firmware compiler for edge AI

Darml — Model in. Firmware out.

A model-to-firmware compiler for edge AI. Upload a trained model (.tflite, .onnx, or scikit-learn .pkl), pick a target hardware, and get a flashable firmware binary or a drop-in C library.

11 supported targets across five hardware tiers — from a 2 KB AVR up to a multi-GB Jetson — through a single uniform pipeline:

parse → check size → quantize → convert → compile → package

Quick start

pip install darml
darml build path/to/model.tflite --target esp32-s3
darml flash darml-<build_id>.zip --port /dev/ttyUSB0

The first build downloads the cross-compiler toolchain (~5 minutes, ~500 MB per platform). Subsequent builds reuse the toolchain cache.

Hardware support

Target	RAM	Flash	Runtime
avr-mega328	2 KB	32 KB	emlearn
avr-mega2560	8 KB	256 KB	emlearn
stm32f4	320 KB	1 MB	TFLite Micro
stm32h7	1 MB	2 MB	TFLite Micro
stm32n6	1.5 MB	4 MB	TFLite Micro
esp32	520 KB	4 MB	TFLite Micro
esp32-s3	512 KB+8 MB	16 MB	TFLite Micro
rpi4 / rpi5	4–8 GB	—	TFLite
jetson-nano	4 GB	—	TensorRT/TFLite
jetson-orin	8 GB	—	TensorRT/TFLite

Run darml targets for the full list with PlatformIO board IDs.

CLI

darml info <model_file>                         # parse + show metadata
darml check <model_file> --target <target>      # estimate fit on hardware
darml targets                                   # list supported hardware
darml build <model_file> --target <target>      # parse → check → compile → zip
darml flash <artifact.zip> --port <serial>      # flash onto a device
darml version                                   # version + license status

darml --help and darml <subcommand> --help for full flags.

Optional dependencies

The base install pulls only what's needed for the CLI to start. Pick extras based on which model formats you'll feed it:

pip install 'darml[onnx]'        # ONNX parsing
pip install 'darml[tflite]'      # TFLite parsing without TensorFlow
pip install 'darml[sklearn]'     # sklearn parsing + emlearn → AVR C
pip install 'darml[build]'       # PlatformIO for MCU firmware compiles
pip install 'darml[all]'         # everything above

Free tier

Darml Core is free under the MIT license, with a soft cap of 5 builds per UTC-day enforced via ~/.darml/counter. The CLI shows your remaining count after each build.

$ darml build model.tflite --target esp32-s3
Build completed: 5f3e…
4 builds remaining today (resets at midnight UTC).

The cap is per-machine and per-user; deleting the counter file resets it. This is friction-as-marketing, not DRM — Darml Pro (which lifts the cap and adds quantization, ONNX conversion, web dashboard, and build cache) is a separate package and a separate purchase.

License

MIT — use Darml Core for anything, commercial or not, no restrictions.