Does the ESP32 firmware call Jev or any cloud API?

No. Training stays in Python, the weights are packed into firmware/main/weights.bin, and the ESP-IDF firmware runs the one-pass scorer entirely on-device in C without cloud calls or GPIO output.

What does the jevlike-esp32 demo model actually decide?

It classifies ASCII input into three routes: command, weather, or complex. On boot the serial console classifies three sample phrases and then reads more lines from UART, printing the same decisions as the Python router.

How do I build and flash the ESP32 firmware?

Build from the example with `idf.py set-target esp32`, then `idf.py build` and `idf.py flash`. Training and weight packing run through examples/esp32_demo/fake_esp32.py. The project targets ESP-IDF and prints sample classifications on the serial console at boot.

Does the firmware need a tensor library or a model download?

No. The TinyScorer weights are packed into firmware/main/weights.bin, and the C scorer implements byte embeddings, layer-norm and matrix multiplies directly. No tensor library, ONNX runtime or network download is linked into the firmware.

Do the Python and ESP32 scorers agree on their answers?

Yes. The README reports that the Python router and the ESP32 firmware print the same route for the boot sample phrases, because the same scorer runs either as Python or as ESP-IDF firmware while training stays in Python.

Is jevlike-esp32 the same thing as TypeSafe's Jev?

No. It follows the separate vinnylarue/jevlike one-pass text model and adds an ESP32 deployment path. The demo does not call a Jev endpoint or use the TypeSafe SDK, and there is no cloud inference.

What license covers it, and can I use it in production?

MIT-licensed and self-contained, but the README states it is intentionally a demo rather than a production model. It classifies ASCII text into command, weather or complex, has no GPIO or weather client yet, and is meant for experimentation.

CommunityJev Models & Reimplementations

Jevlike Esp32

Edge deployment demo that exports a jevlike one-pass text scorer as ESP-IDF firmware; packed TinyScorer weights run on the ESP32 in C with no tensor library and classify ASCII text as command, weather, or complex, alongside a Python trainer and host-side check.

See highlights

Open github.com

Stars: 3
Language: Python
License: MIT
Published by: Community
Last verified: Oct 2026

david-cermak/jevlike-esp32

User repository on GitHub

View on GitHub

Jevlike edge router on ESP32

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Experiments with jevlike ai on embedded devices. Small demo of jevlike repository's text model. Training stays in Python. The same scorer then runs either in Python or as an ESP-IDF firmware.

Quoted from the project's README. Read the full README.

Tagscommunity esp32 embedded firmware

Works forJev for Routing Jev for Classification

Used inJev for Software & DevTools

Languages: Python
Capabilities: Open sourceRuns locally

Highlights

The demo model classifies ASCII text into command, weather, or complex and prints the routing decision; there is no GPIO, weather client, or cloud call yet.
Weights are packed into firmware/main/weights.bin; the C scorer uses byte embeddings, layer-norm, and matrix multiplies, with no extra tensor library.
Training and packing run via examples/esp32_demo/fake_esp32.py, and the firmware builds with idf.py set-target esp32, build, flash.
The Python router and the ESP32 firmware print the same route for the boot sample phrases.
The README states the project is intentionally a demo, not a production model.

Watch out

MIT-licensed and self-contained, but the README states it is intentionally a demo rather than a production model; it follows the separate vinnylarouge/jevlike project's text model and adds an ESP32 deployment path.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

Laya on Hugging Face

The Hugging Face hub for Laya's open System 1 checkpoints: the 421M English model at the repo root, a 322M multilingual build in a subfolder, and the typed-decisions fine-tune, all Apache-2.0.