---
license: mit
library_name: pytorch
pipeline_tag: text-generation
tags:
  - chess
  - transformer
  - quantization
  - gqa
  - rope
---

# ChessNano

A compact autoregressive transformer that reads chess moves in
Standard Algebraic Notation and predicts the next one. No board representation and
no evaluation function — the model sees notation and nothing else.

Code: [github.com/Kcbir/chessnano](https://github.com/Kcbir/chessnano)

## Files

| File | Size | Description |
|---|---|---|
| `chessnano_deployed.pt` | 58.4 MB | Quantised artifact — INT8 weights, bit-packed sign bits, fp16 embedding |
| `chessnano_fp16.pt` | 102 MB | fp16 checkpoint |
| `vocab.json` | 76 KB | SAN token vocabulary |

## Architecture

| | |
|---|---|
| Layers | 8 |
| Model dim | 768 |
| Attention | grouped-query, 12 query heads / 4 KV heads, head dim 64 |
| Positions | RoPE |
| Feedforward | SwiGLU, 2048 hidden |
| Normalisation | RMSNorm, pre-norm, no bias |
| Vocabulary | 2048 SAN tokens |
| Context | 512 tokens |
| Parameters | 51.9M |

## Quantization

Weights are trained with quantization simulated in the forward pass. The
TurboQuant path applies a Hadamard rotation, quantises each coordinate pair's
angle to 128 levels, and stores a one-bit Johnson–Lindenstrauss residual sketch
scaled by the least-squares coefficient. Training and export share one
reconstruction function, so the simulated quantizer matches the exported one.

## Usage

```python
import chess
from inference.tot_inference import ChessNanoPlayer, TotConfig

player = ChessNanoPlayer("chessnano_deployed.pt", tot_cfg=TotConfig(depth=1))

board, history = chess.Board(), []
for san in ["e4", "e5", "Nf3"]:
    board.push_san(san)
    history.append(san)

print(player.predict_move(history, board))
```

Candidate moves are masked against `python-chess`'s legal move list before
sampling, so the decoder cannot emit an illegal move.

## Licence

MIT