Updates

Board editors

Updated models and made quantization more efficient. Optimized the browser engine for much faster analysis.

Y analysis now uses a stronger transformer model. Browser analysis is faster in Firefox.

Improved quantization makes the analysis networks about 15% smaller to download, with similar output quality.

Y and Bent-Y analysis now skips dead and captured cells and recognizes bridge connections when detecting wins.

Analyze positions in your browser with KataHex.

Live analysis: Get live engine analysis as you explore a position. The editors also support branching and drag-to-move editing.
Batch analysis: Batched GPU inference quickly evaluates each position in a game, producing a move-by-move graph.
Networks: The default models were trained from scratch with self-play in a KataGo/KataHex-style pipeline and have about 2 million parameters each. Hex and Y use CNNs; Yavalath and Bent-Y use transformers. Bent-Y uses graph attention to follow the board's geometry. The larger KataHex model is also available for Hex.
In the browser: WebGPU runs the analysis locally, using quantized weights for smaller downloads and FP16 and Winograd convolutions where supported to speed up inference.

Rex (misère Hex) now has an interactive solver for boards through 6×6.

Young and Hayward's Solrex paper established that the first player wins 6×6 Rex.
Drawing on ideas from the paper, the browser solver uses a complete, independently checked decision diagram to look up position outcomes. The 6×6 download is about 3.9 MB.

The second player wins 6×6 Rex+ with optimal play. Rex+ is misère Hex where players place one or more stones per turn.

Veronika Keras's unique-winning-set theorem shows that if any first turn wins, a symmetry-invariant one must win. The proof checks all 2,047 such candidates and rules each out.
The proof and verification code are on GitHub; the verifier recomputes the result without a large proof download or saved search state.
The online solver stops at 5×5: the proof settles the empty 6×6 board, but does not solve every position. An interactive solver would also need to answer arbitrary positions quickly with a practical browser download size.

With optimal play, Red wins standard 4×4 Dark Hex about 79.34% of the time.

Every 4×4 opening is also solved for the variant where Red announces the first move. With the best announced opening, Red's win probability drops to 74.02%.
Exact optimal strategies for both players give matching bounds against unrestricted best responses. The strategies and verification code are on GitHub.