What is Atome LM v2 (SuperESP)?

Atome LM v2, codenamed SuperESP, turns the tiny ternary Atome engine into a suite of on-device AI applications — 11 sensor classifiers plus an on-device dispatcher — that run fully offline on a $5 ESP32 instead of generating text.

Does it run on any ESP32?

Yes. One installer auto-detects the chip (ESP32, S2, S3, C3, C6, H2 — both Xtensa and RISC-V) and flashes the matching prebuilt firmware. The ~27 KB model state fits every mainline variant with large headroom.

Can I train my own classifier without machine-learning experience?

Yes. A logger firmware records your sensor readings to a CSV, then a single train command builds a bit-exact, signed model you flash to the board.

Atome LM v2 (SuperESP): 12 On-Device AI Applications on a $5 Chip

Atome LM started as one tiny ternary language model you could run on a $5 microcontroller. Atome LM v2 — codenamed SuperESP — turns that same 1.58-bit engine into a suite of on-device AI applications: instead of generating text, the chip now classifies the world around it, fully offline.

What v2 adds

12 on-device apps — 11 applied "heads" plus an on-device OS dispatcher: agriculture, voice commands, motion/gesture, sound events, machine anomaly, air quality, energy/NILM, occupancy, wearable activity, water-leak, predictive time-to-failure — each a tiny ternary classifier.
Runs on the real chip: all 12 ran on a physical ESP32-WROOM-32 with ~27 KB of state and 265 KB free heap, every on-device decision matching the host bit-for-bit.
Works on any ESP32: one installer auto-detects the chip (ESP32 / S2 / S3 / C3 / C6 / H2, Xtensa or RISC-V) and flashes the matching firmware — no toolchain to install.
Make your own in minutes: a logger firmware records your sensor to CSV, then train → report → flash builds a custom classifier. No ML expertise.
Trust built in: every model is Ed25519-signed and integrity-checked on load, with a tamper-evident log of decisions — auditable edge AI.

Honest about what it is

This is a real, production-grade open kit — not magic. Several demo heads ship on physics-grounded synthetic data (clearly labelled), and you replace them with your own real data through the same CSV path. Voice keyword-spotting works but is modest on a 20 KB ternary model. We publish the numbers, the tests, and the honest limits rather than a marketing benchmark.

Get it

The v1 engine and weights are already public on GitHub and Hugging Face under Apache-2.0. The v2 SuperESP kit ships the applied layer on top. If you'd rather have it done, certified, or licensed for your product, see services.