Skip to content
HardwareLab
Local LLM Rigs & Edge AgentsAdvanced

PFor

$6.80 (Wireless-Tag store)A 180.9M-parameter MoE language model and agent on a $6.80 ESP32-P4.

Full build guide

PFor
No photo in the source; GitHub repository card, served from HardwareLab.

By cyfrit

View original on GitHub
Posted
GitHub stars
8 (GitHub)

Build sheet

Board
WT9932P4-Tiny (ESP32-P4 v1.3, 32MB PSRAM, 16MB flash)
AI model or service
PFor: 12-layer MoE (29 experts, top-1) + PLE, 1.58-bit ternary weights, Qwen3.5-pruned vocab
Difficulty
Advanced
Est. cost
$6.80 (Wireless-Tag store)

Overview

Custom 180.9M-parameter MoE+PLE model quantized to W1.58A8 (~44 MiB across PSRAM and XIP flash) runs ChatML chat and a fixed-tool agent loop at ~9 tok/s on a WT9932P4-Tiny ESP32-P4 board using Espressif's XespV acceleration.

How it works

A 12-layer MoE with Per-Layer Embeddings and 1.58-bit weights lives across PSRAM and flash, running chat and a fixed-tool agent loop at about 9 tok/s.

Hardware

  • WT9932P4-Tiny (ESP32-P4 v1.3, 32MB PSRAM, 16MB flash)

More ESP32 builds

Like this build? Get the weekly builds

Get the best new AI hardware builds every Thursday. One email a week, free.

Built something? Submit a build

This page is a short summary written by HardwareLab. Build steps, code and photos belong to the creator; follow “View original” for the full project.

Featured? Add this badge

Built this? Paste the badge into your README. It links back to this page.

Featured on HardwareLab

Markdown

[![Featured on HardwareLab](https://hardwarelab.dev/badge.svg)](https://hardwarelab.dev/projects/pfor-esp32-p4/?utm_source=badge&utm_medium=referral&utm_campaign=launch_2w)

HTML

<a href="https://hardwarelab.dev/projects/pfor-esp32-p4/?utm_source=badge&utm_medium=referral&utm_campaign=launch_2w"><img src="https://hardwarelab.dev/badge.svg" alt="Featured on HardwareLab" width="188" height="20"></a>

Related builds

All 22 in Local LLM Rigs & Edge Agents