Luminex NNUE
A tiny neural network for chess position evaluation, trained for the Luminex chess engine.
Model Details
| Property | Value |
|---|---|
| Architecture | 768 → 32 → 1 |
| Parameters | 24,641 |
| File size | 48.2 KB (int16 quantized) |
| Activation | Clipped ReLU (hidden), Sigmoid (output) |
| Input | 768 binary features (12 piece types × 64 squares) |
| Output | Win probability [0, 1] |
Training
| Property | Value |
|---|---|
| Data | 1,000,000 real Stockfish-evaluated positions |
| Source | mateuszgrzyb/lichess-stockfish-normalized |
| Epochs | 100 |
| Best val loss | 0.034940 (MSE) |
| Optimizer | Adam (lr=0.001, cosine annealing) |
| Hardware | Google Colab free T4 GPU |
| Training time | ~20 minutes |
Files
luminex-nnue-32.bin— Quantized int16 weights for C++ engine inferenceluminex-nnue-32.pt— PyTorch checkpoint for retraining
Binary Format
5 bytes — magic "LNNUE"
int32 — version (1)
int32 — input size (768)
int32 — hidden size (32)
int32 — L1_SCALE (255)
int32 — OUT_SCALE (64)
int16[32×768] — feature transformer weights
int16[32] — feature transformer biases
int16[1×32] — output weights
int16[1] — output bias
Usage
Place luminex-nnue-32.bin next to the Luminex executable. The engine loads it at startup automatically.
License
MIT
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support