An ESP32 Speaks, Chats, and Gets a New $21 Brain in Today's All-Microcontroller AI Special

A 3.8MB Voice Model Speaks Straight From an ESP32-S3

Developer lokutor-ai pushed ito to GitHub, a streaming text-to-speech model built for the ESP32-S3. The model has 3.3 million parameters and ships as 3.8 MB of weights, small enough to sit in flash alongside the rest of a firmware image. The repo describes natural-sounding speech generated with no cloud call, created on October 2, 2026 and now circulating on GitHub's trending feed.

The project is young — 32 stars, no published benchmark against commercial TTS engines, and the README itself frames this as an early build rather than a finished product. There is no public audio comparison yet against cloud TTS quality, so treat the "natural-sounding" claim as the developer's own description until independent samples surface.

What matters is the size. A sub-4 MB voice model that runs fully on-chip keeps joining a growing list of offline voice pipelines on ESP32-class silicon, pushing the question of whether a smart speaker needs a server at all further toward "no."

An Early ChatML Agent Runs Directly on the ESP32-P4

Developer cyfrit published p-for-llm (PFor), a small language model running on the ESP32-P4. The repo describes Instruct (ChatML) and Agent modes, with inference running at roughly 9 tokens per second on the chip's RISC-V core.

The developer's own description calls both capabilities "extremely early and highly unstable," which is an honest flag worth repeating: this is a proof that a ChatML-style chat format and basic tool-calling structure can execute on microcontroller-class hardware, not a usable assistant yet. No benchmark numbers or sample transcripts beyond the author's note have been published.

Even so, the framing matters more than the polish. Pairing an agent-style interface with a sub-$15 chip keeps nudging the floor for "where can a language model physically run" lower, following the same trajectory as other ESP32 LLM experiments documented this year.

A $21 Board Pairs an ESP32-P4X With a WiFi 6 Co-Processor

CNX Software reported on November 4, 2026 that Wireless Tag's WT99P461-S1 board combines an ESP32-P4X with an ESP32-C61 on the same WT01P461-S1 module first seen on the ESP32P4C61-TINY board last June. The board is priced at $21 and adds Ethernet, WiFi 6, RS485, CAN bus, and MIPI CSI/DSI camera and display connectors.

The listed price and spec sheet come from the vendor's own board page as relayed by CNX Software; actual retail availability and shipping region were not confirmed in the report, so treat the $21 figure as the manufacturer's quoted price rather than a confirmed street price.

The combination is the interesting part: the P4X brings more compute headroom for the kind of small models and vision pipelines running on ESP32-P4 boards elsewhere in this dispatch, while the C61 co-processor handles WiFi 6 radio duty efficiently — a cheap, full-featured base for builders who want camera input and a chat-capable model on the same board.

Three ESP32 stories, three different jobs: speaking, chatting, and simply having enough radio and compute to try either. None of it is finished software, but the chip keeps absorbing more of the stack it used to need a server for.