---
language: en
license: mit
tags:
- chess
- transformer
- custom-architecture
pipeline_tag: robotics
---

# ChessDumb-2.5M (`dumbc`)

**ChessDumb-2.5M** is a hyper-expressive, ultra-compact Chess AI architecture designed for extreme representation capacity with only **1.69M parameters**.

## Key Features
- **DumbChessRetina**: Non-Euclidean grid folding with 4-axis directional convolutions and ray-tracing threat embeddings.
- **DumbAttention**: 4-way fused attention featuring Hadamard tensor braids, 3-body trinity tensor coupling, material differential biases, and dynamic temperature Softmax.
- **DumbFractalFFN**: Channel-shuffled polynomial interaction FFN.
- **DumbRecurrentEngine**: Virtual 15-layer deep thinking engine powered by a 5-layer parameter-reusing loop with step embeddings.

## Usage
```python
import torch
from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("YOUR_USERNAME/ChessDumb", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("YOUR_USERNAME/ChessDumb", trust_remote_code=True)

# Dummy Forward Test
board = torch.randint(0, 14, (1, 64))
mat_diff = torch.randn(1, 64, 64)
mat_weights = torch.randn(1, 64)

outputs = model(board_state=board, mat_diff_matrix=mat_diff, material_weights=mat_weights)
print("Policy shape:", outputs["policy_matrix"].shape) # (1, 64, 64)
print("Value shape:", outputs["value_logits"].shape)   # (1, 3)

```