mateuszgrzyb/searchless-chess-vit-medium

♟️ Searchless Chess: ViT-Medium (1960 ELO)

This is the flagship model of the Searchless Chess project—a neural network that plays chess at a master level (~1960 ELO) relying purely on neural intuition, without any search algorithms (no minimax, no alpha-beta pruning).

• Architecture: Vision Transformer (ViT)
• Parameters: 9.5M
• Estimated ELO: 1960
• Training Data: 316M positions from Lichess/Stockfish

🎯 Key Achievements

• Master-Level Play: Reached ~1960 ELO by evaluating only the current board state.
• Extreme Efficiency: Achieved 67% of DeepMind's searchless ELO (2895) using less than 4% of their parameters and training data.
• ViT Superiority: This Medium ViT model outperforms ResNet architectures that are 2.5x larger, proving that Transformers are exceptionally efficient at capturing global board dependencies.

📊 Performance Metrics

Metric	Value
ELO (Estimated)	1960
Tier 4 Puzzles (1000-1250)	88% Accuracy
Tier 6 Puzzles (1500-1750)	86% Accuracy
Tier 8 Puzzles (2000-2250)	45% Accuracy

🚀 Quick Start & Usage

This repository provides everything needed to run or fine-tune the model:

• model.keras: Complete model (architecture + weights) in Keras 3 format.
• model.weights.h5: Standalone weights.
• config.json: Model architecture configuration.

Which file to use?

• Use model.keras for a quick "plug-and-play" experience, as it contains both the model architecture and the trained weights in one file.
• Use config.json and model.weights.h5 if you prefer to build the model structure manually in your code and load only the parameters (e.g., for custom fine-tuning).

Using the model in Python

To use this model for playing chess, use the ChessAI wrapper provided in the official GitHub repository. It handles the FEN-to-Tensor encoding and move selection logic.

# 1. Clone the repository: git clone https://github.com/mateuszgrzyb-pl/searchless-chess
# 2. Use the following code:

import chess
from src.chess_ai.core.model import ChessAI

# Load the chess engine (provide path to the downloaded model.keras)
chess_bot = ChessAI('path/to/vit-medium/model.keras')

# Create a chess board
board = chess.Board()

# Play a game
board.push_san("e4")  # Your move
engine_move = chess_bot.make_move(board)  # Engine responds via neural intuition
board.push(engine_move)

print(board)
print(f"Engine played: {engine_move}")

For detailed instructions on board encoding, training, and evaluation, visit the Searchless Chess GitHub.

🔬 Model Architecture

The model treats the chess board as a set of tokens. Unlike traditional CNNs that look at local patterns, the Vision Transformer uses self-attention to understand the global relationship between all pieces simultaneously (e.g., a bishop's long-range pressure on a king).

• Patch Size: 1x1 (each square is a token)
• Embedding Dim: 384
• Num of Heads: 6
• Transformer Layers: 8

📂 Dataset

The model was trained on the Lichess Stockfish Normalized dataset, consisting of 316M deduplicated positions with evaluations.

📜 Citation

If you use this model in your research, please cite:

@software{grzyb2025searchless,
  author = {Grzyb, Mateusz},
  title = {Searchless Chess: Master-Level Chess Through Pure Neural Intuition},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/mateuszgrzyb/searchless-chess-vit-medium}
}

📧 Contact

• Project Repository: GitHub
• LinkedIn: Mateusz Grzyb
• Website: MateuszGrzyb.pl