Stress-testing Gemini
Deep-Formix started as my attempt to stress-test Gemini, which I considered the best AI model at the time. It made a lot of bugs, but I still managed to get a working version of depth-four full search on the GPU, using the full chess rules.
That was the practical goal: get the engine working despite the bugs. The project also brought two existing GPU chess approaches together and added a new evaluation architecture.
The biggest achievement for me was the custom neural evaluation network. I wanted it to make complex decisions directly from a position, using a long tail of small, fully connected 32×32 layers: little computation at each layer, but strong expressive power through depth. The architecture also supports the NNUE approach to efficient updates.
The work investigates how much chess evaluation can be carried by a learned model rather than a long search. It combines C++/CUDA engine code with Python training tools.
The main idea
The README describes a 12-layer NNUE architecture with 32×32 layers and skip connections. It uses 32 model buckets selected by piece count, pawn count, and the presence of queens.
Instead of using one evaluation model across all positions, the bucket scheme chooses a model suited to the material configuration. The project is inspired by searchless chess, while the engine also includes search code.
Implementation
The move generator refactors GPU Perft, while the engine structure and minimax implementation draw on Chess-Engine-GPU. The project brings those components together and adds its deep evaluation architecture, new evaluation and position classes, and model training and conversion pipeline. The repository documents a UCI interface, PyTorch training and testing scripts, and conversion to the binary model used by C++.
The README identifies the upstream sources. Those links matter for understanding both the architecture and its provenance.
Reported results
The author reports that searchless evaluation plays within approximately 40 Elo of Stockfish at one ply. That is a repository-reported comparison, not an independently reproduced benchmark by this website. It should not be read as a claim of parity with full-strength Stockfish.
The documentation also reports training on a searchless dataset and a multi-day training run on an RTX 3080 Ti. The practical result here is an engine and training experiment; strength depends on the model, hardware, and evaluation setup.
Inspect the experiment
Start with the source and training documentation. There is no verified hosted demo linked in the README. The repository also flags uncertainty around the licensing of derivative portions, which should be checked there before redistribution.
