ESP32 Special: Small LLMs Learn to Chat, Listen and Keep the Fish Alive

A Cardputer Learns to Hold a Conversation With 512KB of RAM

Builder therezor has an 8-million-parameter chatbot running entirely offline on the ESP32-S3 Cardputer ADV. The Q4 weights are baked straight into firmware, no SD card or external flash chip needed, and it produces roughly 5 tokens per second inside the board's 512KB of RAM.

The project claims this is the first chatbot to hold a coherent English conversation on a microcontroller this small. That claim is self-reported and 'coherent' likely means short, toy-scale exchanges rather than sustained dialogue — worth treating as a proof of concept, not a Siri replacement.

Still, packing conversational behavior into 512KB with zero extra memory is a meaningful shrink from the flash-hungry LLM demos that have dominated this beat, useful for badge-class and wearable ESP32 hardware.

A $5 Microcontroller Chip Claims It Beats Whisper-Tiny

The Lokutor team posted Oído, a speech-recognition model built on an int8 NVIDIA Conformer-CTC Small (13M parameters) that runs on an ESP32-S3 with 8MB PSRAM — no GPU, no NPU, no cloud round trip.

The team's own benchmark claims it beats Whisper-tiny, but that comparison comes from the builders themselves, without a public dataset breakdown or independent replication yet, so treat the win as self-reported until others test it.

Even discounted, running a Conformer-style ASR model on stock ESP32 silicon rather than a keyword spotter is a real step for offline voice pipelines on cheap boards.

A Distilled 14M-Parameter LLM Keeps a Virtual Fish Tank Alive

pocket-tank, built by mediacutlet, distills a 26B-parameter teacher model down to 14M parameters to drive fish behavior on an $8 ESP32-S3. Four-bit quantization shrinks it from 57MB in FP32 to 7.56MB, and memory-mapping weights straight from flash lets the board's PSRAM handle inference without copying everything into RAM.

On real hardware it generates about 12 tokens per second, with each fish decision taking roughly 3.7 seconds, while the tank animation keeps rendering at 25-30fps on the other CPU core, per the project's own numbers and a companion write-up on Electronics For U.

It's a toy, not a chatbot — but it's a clean example of distillation making a real transformer cheap enough to run a game-logic loop instead of a scripted state machine.

A 30.7M-Parameter Dense LLM Squeezes Onto an $8 ESP32-S3

Developer JARACH-209 published esp32-30.7M, a dense 30.72-million-parameter language model quantized to 4 bits and run on a single ESP32-S3, with no cloud connection and no SD card.

The honest caveat is right in the repo: it runs at 0.95 tokens per second, under one word a second, which makes it a proof of scale rather than something usable for a real conversation.

It's the largest dense model shown running natively on ESP32-class hardware this week, and it fits the pattern this whole beat keeps testing: how far the same $8 chip can be pushed before speed collapses entirely.

Waveshare's New ESP32-C5 Devkit Adds a 3.5-Inch Touchscreen

Waveshare's ESP32-C5-Touch-LCD-3.5, covered by CNX Software on September 30, 2026, pairs Espressif's dual-band ESP32-C5 with a 3.5-inch color touch LCD, plus optional BF3901 camera and battery add-ons.

It's a devkit, not a finished product, so real-world power draw and camera performance still depend on what the buyer bolts on. But a Wi-Fi 6-capable ESP32 variant with a built-in touch display in this size class gives builders a ready base for local vision or voice UI experiments without designing a carrier board from scratch.

A Student Team's ESP32-S3 Soccer Robots Finish 11th at RoboCup Junior

Espressif's own developer blog posted a build log on September 30, 2026, from Team XLC-WYLDFYRE, who built two autonomous soccer robots around the ESP32-S3 and finished 11th of 24 teams at RoboCup Junior 2026 in Incheon.

This isn't a model-on-device story — it's a real maker build with a dated competition result straight from the vendor's own blog, showing the ESP32-S3 doing sensor fusion and control loops for autonomous soccer play rather than any language model.

Worth including on the beat because it's exactly the kind of hands-on, student-built firmware work that makes ESP32-class silicon a teaching platform, not just a chatbot host.

Six stories, one chip family: the same ESP32-S3 running chatbots, speech recognition, fish-tank logic and soccer robots this week, alongside a new devkit from Waveshare. Every claim above is self-reported by its builders, so read the repos before trusting the tokens-per-second numbers.