# HN: I ran a 56M parameter LLM across three ESP32-S3 boards via ESP-NOW > Source: > Published: 2026-07-30 15:07:32+00:00 **Distributed micro-LLM inference across three ESP32-S3 N16R8 boards with ESP-NOW communication.** This project implements a **distributed AI system** that runs a **56M-parameter language model** across three ESP32-S3 microcontrollers. Inspired by [slvDev/esp32-ai](https://github.com/slvDev/esp32-ai), which demonstrated running TinyStories on a single board, this project extends the architecture to a multi-board distributed system with web-based interaction. The model is trained on **WikiText-103** (Wikipedia corpus) using **Per-Layer Embeddings (PLE)** from Google's Gemma architecture, quantized to 4-bit, and split across three boards that communicate via ESP-NOW wireless protocol. The 50.3M-parameter PLE table is split across Board A and Board B (**Split-PLE**) to fit within the 16MB flash per board. Board B maintains a **KV cache** in PSRAM (1.5 MB for 256 positions), enabling the transformer to attend to the full generated sequence instead of operating token-by-token. ``` ┌───────────────────┐ ESP-NOW ┌───────────────────┐ ESP-NOW ┌───────────────────┐ │ Board A │ ◄──────────────► │ Board B │ ◄──────────────► │ Board C │ │ Embeddings + │ │ Core + KV Cache │ │ PLE_A + Decoder │ │ PLE_Proj + Head │ │ │ │ + WiFi │ │ │ │ │ │ │ │ • BPE Tokenizer │ │ • 6 Attn layers │ │ • PLE table A │ │ • tok_emb 8-bit │ │ • 6 FFN layers │ │ (12.5 MB) │ │ • ple_model_proj │ │ • KV Cache (128) │ │ • Sampling │ │ • out_norm + head │ │ • PLE_B (half) │ │ • Web Server │ │ │ │ • PLE Gating │ │ • Space heuristic │ │ Flash: 4.31 MB │ │ Flash: 13.71 MB │ │ Flash: 13.37 MB │ └───────────────────┘ └───────────────────┘ └───────────────────┘ Browser ──WiFi──► Board C ──ESP-NOW──► Board A ──ESP-NOW──► Board B ◄─────────────── ◄────────────── ``` **User** types a prompt in the browser (connected to Board C's WiFi AP) and clicks "Generate"**Board C** receives the prompt via HTTP POST → forwards it to Board A via ESP-NOW (`MSG_EMBED_REQUEST` )**Board A** tokenizes the prompt text → for each token: a. Looks up token embedding`x[D]` (8-bit tok_emb) b. Computes local PLE projection:`tmpP = ple_model_proj_a @ x` , RMS-norms it c. Requests PLE table row from**Board C** via ESP-NOW: sends`token_id` → C looks up`ple_table_a[token]` → sends row back d. Combines:`ple = (tmpP + trow * sqrt(Ph)) / sqrt(2)` → sends`token_id + x[D] + ple[L×Ph]` to**Board B** e.**Board B** computes PLE_B[L×Ph] from its local table → combines PLE_A + PLE_B into full PLE[L×P] → runs 6 transformer layers (attention + FFN + PLE gating) → sends hidden state`x[D]` back to Board A f.**Board A** applies`out_norm` → computes output head:`logits[V] = tok_emb^T · rmsnorm(x)` → sends top-40 token IDs to Board C g.**Board C** samples the next token → decodes via BPE vocab → appends to generated output**Board C** streams the generated text to the browser via SSE (`/stream` endpoint)**User** sees the text appear in real-time on the web page The 50.3M-parameter PLE table (vocab × layers × 256) is split into two 128-dimension halves: **PLE_A**(Board C, 12.5 MB, 4-bit group=64): first 128 dims per layer — reloaded from Board C at runtime via ESP-NOW (MSG_PLE_REQUEST / MSG_PLE_RESULT)**PLE_B**(Board B, 12.6 MB, 4-bit group=128): last 128 dims per layer — local on Board B Board A keeps the small `ple_model_proj` (~50 KB, 4-bit) and `ple_proj_norm` (~0.5 KB, fp32) for local projection. The 12.5 MB PLE table was moved from Board A to Board C to fix partition overflow on Board A and allow upgrading `tok_emb` to 8-bit for better inference quality. Each token in the autoregressive loop previously ran through Board B's transformer with `seq_len=1` — attention only saw the current token, not the history. This crippled coherence because the model was designed for `seq_len=128` context. Board B now maintains a **KV cache** in PSRAM (1.5 MB for 256 positions): - For each new token, Q is computed normally, while K and V are appended to per-layer caches - Attention scores are computed against **all cached positions**(not just the current token) - RoPE uses the actual position index instead of `pos=0` — for the first time the transformer sees proper position-aware attention - Cache is reset on `MSG_SEQ_START` (new prompt) and caps at`seq_len=256` - No changes to A, C, or the wireless protocol This is the single largest quality improvement: the transformer finally works as designed, attending to the full generated sequence instead of operating token-by-token in isolation. | Board | MAC Address | Serial Port | |---|---|---| | A (Embeddings + PLE_Proj + Head) | `14:c1:9f:2a:ac:c8` | `/dev/cu.usbmodem5C372059631` | | B (Core + KV Cache) | `14:c1:9f:2c:91:10` | `/dev/cu.usbmodem5C372065471` | | C (PLE_A + Decoder + WiFi AP) | `28:84:85:51:dc:10` | `/dev/cu.usbmodem5C4D0363671` | **ESP-NOW** wireless peer-to-peer (no router needed)- Packets ≤ 250 bytes with automatic fragmentation - Custom protocol with sequence numbers and acknowledgments - Estimated latency: 1–5ms per round-trip ``` esp32s3/ ├── README.md # This file (English) ├── README.es.md # Spanish version ├── pyproject.toml # Python dependencies │ ├── src/ # Training pipeline (runs on PC/Mac) │ ├── model.py # PLE TinyLM architecture (PyTorch) │ ├── dataset.py # TinyStories/WikiText-103 data loading │ ├── train.py # Training loop │ ├── quantize.py # 4-bit group-wise quantization │ ├── export.py # Export to 3 board binaries │ └── gen_assets.py # Generate vocab.h for firmware │ ├── firmware/ # ESP32-S3 firmware (Arduino IDE) │ ├── common/ │ │ ├── llm.h # Distributed C inference runtime │ │ └── espnow_protocol.h # ESP-NOW message protocol │ │ │ ├── board_a_embeddings/ # Board A firmware │ │ ├── board_a_embeddings.ino # Main sketch │ │ ├── partitions.csv # Flash partition table │ │ └── vocab.h # Generated tokenizer vocabulary │ │ │ ├── board_b_core/ # Board B firmware │ │ ├── board_b_core.ino # Main sketch │ │ └── partitions.csv │ │ │ ├── board_c_decoder/ # Board C firmware │ │ ├── board_c_decoder.ino # Main sketch (WiFi + Web UI) │ │ ├── partitions.csv │ │ └── vocab.h # Generated tokenizer vocabulary │ │ │ └── model/ # Exported model binaries (gitignored) │ ├── board_a.bin # 4.31 MB (tok_emb 8-bit + ple_proj + out_norm) │ ├── board_b.bin # 13.71 MB (transformer layers + PLE_B) │ ├── board_c.bin # 13.37 MB (PLE table A, 4-bit group=64) │ ├── golden.npz # Reference logits for verification │ └── golden.txt # Text-format golden reference │ ├── tools/ # Utility scripts │ ├── flash_all.sh # Flash all 3 boards │ ├── verify_models.py # Verify exported binaries │ └── setup_env.sh # Install Python dependencies │ ├── data/ # Dataset (gitignored) │ ├── tinystories/ # Raw TinyStories text │ ├── wikitext103/ # Raw WikiText-103 text │ ├── tokenizer/ # Trained BPE tokenizer │ ├── train.bin # Tokenized training data (112M tokens) │ └── val.bin # Tokenized validation data │ └── runs/ # Training checkpoints (gitignored) ├── ple-wiki60m-s0.pt # 56M model checkpoint ├── ple-wiki60m-s0.json # Training history └── train.log # Training output log ``` | Parameter | Value | |---|---|---| | Architecture | Tiny decoder-only transformer with Split-PLE | | Total Parameters | 56.0M stored | | Core (dense, SRAM) | 1.5M | | PLE Table (flash, split) | 50.3M (25.2M per board) | | Output Head (tied) | 4.2M | | Vocabulary Size | 32,768 (BPE, WikiText-103) | | d_model | 128 | | n_layers | 6 | | n_heads | 4 | | ffn_hidden | 223 | | ple_dim | 256 (128 per board) | | Quantization | tok_emb: 8-bit group=128, PLE table A: 4-bit group=64, rest: 4-bit group=128 | | Model Binary Size | 31.39 MB total (A: 4.31, B: 13.71, C: 13.37) | | Dataset | WikiText-103 (Wikipedia) | | Training Steps | 12,000 | | Batch Size | 8 | | Sequence Length | 128 | | Metric | Value | |---|---| | Final Validation Loss (fp32) | 5.26 | | Perplexity (fp32) | 192 | | 4-bit Quantization Degradation | +0.62 nats | | 4-bit Perplexity | 358 | | Training Tokens | 12.3M | | Training Time | ~6.9 hours (Mac i7 CPU) | - Python 3.11+ [uv](https://docs.astral.sh/uv/)package manager- Arduino IDE with ESP32 board package - 3x ESP32-S3 N16R8 boards ``` source .venv/bin/activate python src/dataset.py --dataset wikitext103 python src/train.py --arm ple --d-model 128 --n-layers 6 --ple-dim 256 \ --target-core 1500000 --batch-size 8 --seq-len 128 --steps 12000 --tag wiki60m python src/quantize.py --tag wiki60m python src/export.py ple-wiki60m-s0 python src/gen_assets.py # Connect each board one at a time ./tools/flash_all.sh /dev/cu.usbmodemXXXX ``` Note:All 3 boards are already flashed. If you need to reflash, see`tools/flash_all.sh` . - Power on all three boards - Connect your phone/laptop to WiFi: **ESP32-DIST-AI**(password:`ai123456` ) - Open `http://192.168.4.1` in your browser - Type a prompt and click "Generate" ## Architecture — PLE TinyLM v2, Split-PLE, 56M params - PLE TinyLM v2 with d_model=128, n_layers=6, ple_dim=256 - Split-PLE: PLE table divided across Board A + Board C (128+128 per layer) - PLE table A moved from Board A to Board C (fixed partition overflow on A) - Per-token A↔C PLE request/response via ESP-NOW - KV cache on Board B: full multi-head causal attention with K/V caching in PSRAM (1.5 MB for 256 positions), RoPE uses actual position indices - MAX_SEQ_LEN=256, gen loop=128, supporting 30+ word sequences - srand(esp_random()) for non-deterministic sampling ## Training — WikiText-103, PPL 192, 12K steps, 12.3M tokens - WikiText-103 dataset download and BPE tokenization (112M tokens) - Model training: 12,000 steps, batch_size=8, seg_len=128, 12.3M tokens - Final fp32 loss: 5.26, Perplexity: 192 - Training time: ~6.9 hours (Mac i7 CPU) ## Quantization & Export — 4-bit, 31.39 MB total - 4-bit group-wise quantization (+0.62 nats degradation from fp32, 4-bit PPL 358) - tok_emb: 8-bit group=128, PLE table A: 4-bit group=64, rest: 4-bit group=128 - Export to 3 board binaries: A: 4.31 MB, B: 13.71 MB, C: 13.37 MB - Reference golden logits for verification ## Firmware — 3 boards, ESP-NOW, Web UI, autoregressive loop - Board A: tokenizer + 8-bit tok_emb + PLE projection + output head - Board B: 6 transformer layers + PLE_B + PLE gating + KV cache - Board C: PLE table A (4-bit group=64) + decoder + WiFi AP + Web UI - ESP-NOW protocol with fragmentation, sequence numbers, and acks - Web UI for prompt input, streaming output, and SSE endpoint - Autoregressive generation loop (up to 128 tokens) across all 3 boards - Space-insertion heuristic on Board C - Arduino core 3.3.11 compatibility, partition table 1.25 MB ## Flash & Deploy — 3 boards flashed with firmware + model - Board A: MAC `14:c1:9f:2a:ac:c8` / port`5C372059631` - Board B: MAC `14:c1:9f:2c:91:10` / port`5C372065471` - Board C: MAC `28:84:85:51:dc:10` / port`5C4D0363671` - Flash and verification scripts ## Testing — Local inference confirmed coherent output (~30 words) - Local inference test confirmed ~30 coherent words - Confirmed 4-bit quantization is the quality bottleneck (not protocol) - Comparison between full fp32, quantized, and on-device inference - **BPE tokenizer in C**— Replace brute-force vocab search with proper BPE merge implementation - **Static ESP-NOW MACs**— Configure known MACs instead of broadcast - **Web UI improvements**— Streaming token display, temperature/top-k controls, status indicators - **Error handling**— ESP-NOW retransmission and timeout on packet loss - **Power optimization**— Deep sleep between inference requests - **Audio output**— I2S speaker integration (future) | Component | Quantity | Notes | |---|---|---| | ESP32-S3 N16R8 | 3 | 512KB SRAM, 8MB PSRAM, 16MB flash | | USB-C cables | 3 | For flashing firmware | | Computer | 1 | Mac/Linux/Windows with Arduino IDE | This project extends the work of: — Running 28.9M parameter LLM on a single ESP32-S3 using Per-Layer Embeddings[slvDev/esp32-ai](https://github.com/slvDev/esp32-ai)**Google Gemma**— Per-Layer Embeddings architecture** WikiText-103**— Dataset by Salesforce Research ([arXiv:1609.07843](https://arxiv.org/abs/1609.07843))** TinyStories**— Dataset by Eldan & Li (Microsoft Research,[arXiv:2305.07759](https://arxiv.org/abs/2305.07759))— Inspiration for running tiny LMs in plain C[karpathy/llama2.c](https://github.com/karpathy/llama2.c) MIT