Cubiq

The problem

Most serious speedcubers still time their solves in csTimer. It works, but it looks and feels like 2008, and it stops at timing. There was no modern tool that combined a pleasant training screen with proper statistics and help with solving.

I wanted one place to train, track progress, and ask "how should this scramble have been solved?" for any puzzle, from the 2×2 to the Square-1. Cubiq is built for intermediate and advanced cubers who time themselves on several puzzles and care about their averages.

What I built

A timer that stays exact

The clock reads the browser's high-resolution time on every frame, so a long solve never drifts. A 300-millisecond hold arms the timer, the same way competition inspection works. Averages follow the World Cube Association rules, from ao5 up to ao100.

A separate solver for each puzzle

I wrote a solver for each of the eight major puzzles and picked the strategy by how many positions the puzzle has.

• 2×2, Pyraminx and Skewb are small enough to precompute every position; the solver looks up the shortest solution, never longer than 11 moves.
• 3×3 is solved the way people solve it (CFOP): an optimal cross, the first two layers paired, every last-layer case recognised from a full algorithm set, with wasted moves between stages cancelled.
• 4×4 and 5×5 reduce to a 3×3 by solving the centres and pairing the edges, with parity fixes, then finish through the 3×3 solver.
• Megaminx solves layer by layer, with a small library of last-layer sequences.
• Square-1 needs two phases, shape then pieces, with a 3D view that animates its half-turn flips as solid geometry.

The whole app moved into the browser

Every solver first ran through a Python service the user had to start locally, so the Solve Lab simply didn't work unless that service was running. I ported every solver into TypeScript running in a Web Worker, so all eight now solve client-side with no backend involved. The Python service still exists behind the optional research tab, now as the reference implementation the browser ports are checked against. With nothing left needing a server, I shipped Cubiq as a static export to GitHub Pages instead of a hosted Next.js server.

Four sections built around what a cuber is doing

The app began as four separate pages named after their implementation: timer, stats, history, solvers. That split stats from history, two things a cuber wants side by side, and hid the solver suite behind a name that describes code, not a question. I merged stats and history into one Progress page and renamed the solver suite Solve Lab, still one tab per puzzle carrying a scramble over when you switch. A new Learn section adds a spaced-repetition trainer for OLL and PLL cases, and old URLs still redirect so saved links keep working.

A model that teaches itself to solve the cube

A separate Python service trains a neural network the way DeepCubeA did. It scrambles a solved cube, then learns to estimate how many moves it is from solved. A research dashboard shows training live. You can run the model on a scramble and compare its answer with the optimal one from Kociemba's solver, which I use only as a benchmark.

Result

Cubiq has been live and free to use since September 2026. It solves all eight puzzles with solvers I wrote myself, entirely in the browser. I used cubing.js only to draw the cube and scramble it, and Kociemba's solver only to measure against. Every solve is saved in the browser and can be exported. The first plan only had a cross solver, marked as possibly too hard for an early phase. I built the first version over about 11 weeks, from May to July 2026. I then returned to it through September to move every solver client-side, redesign the app around four task-based sections, and ship it as a static site with no server to keep running.