How Jelmata's Elite AI Plays

Jelmata has an “Elite” difficulty setting. It’s a small neural network that fits inside the app bundle, plays instantly, and never calls a server. This post is a tour of what that actually means — why Elite moves feel the way they do, why you never see a “thinking…” spinner, and why you can play it with your phone in airplane mode.

The whole of Elite — every position it’s ever learned from, every instinct it plays with — is a single 303 KB file that shipped with the app you installed from the App Store. That’s smaller than a single screenshot on the store listing. It’s already on your phone the moment the install finishes.

No account, no subscription, no sign-in. If you’ve paid to remove ads or bought any other upgrade, that’s a one-time purchase on your Apple or Google account; the AI itself isn’t something we charge for, and never will be.

When you play against Elite, the response to your move arrives about as fast as you can take your finger off the screen. That’s not because we’re rushing it — it’s because the AI doesn’t need to pause and think.

Most game AIs that play well think by simulating moves and counter-moves, evaluating positions, backing up values. That takes time on every turn, which is why a lot of mobile games make you stare at a spinner or pad the delay so the AI “feels” like it’s deliberating. Jelmata’s Elite does its thinking differently. The thinking happened once, on a computer much more powerful than your phone, months before you ever opened the app. What ships with the game is the trained result — the taste — not the process.

A small neural network trained from nothing would play erratically. Jelmata’s Elite doesn’t, and the reason is the training recipe. Elite was trained to imitate a much stronger program — one that did use lots of compute per move, did search many positions, and did play itself for thousands of games to sharpen its style. We then taught Elite to reproduce that program’s move choices.

The practical effect is that Elite looks like it’s playing the long game even though it’s only thinking one move ahead. It sets up threats several turns in the future. It doesn’t fall for cheap traps. It will exploit a sloppy corner or a weak edge if you give it one. The tiny network has inherited the taste of a big one, the same way a student can absorb a teacher’s aesthetic without reinventing the curriculum.

We designed the AI around a simple test: Jelmata has to feel instant on a budget Android phone from several years ago, with no specialized graphics hardware and a battery that’s been through a few charge cycles. That constraint is why Elite is small, why it doesn’t search, and why it doesn’t drain your battery while you play.

If you’re on a current flagship, Elite is effectively free — you won’t see any difference in performance between it and the simpler difficulty tiers. On older devices, it still runs cleanly. The AI you’re playing is the same on every phone, and it has the same strength regardless of what you’re holding.

Jelmata’s other difficulty tiers — Easy, Medium, Hard — are built on a handful of hand-written scoring rules. They look at the board, count a few well-chosen signals (how many cells this move captures, how it affects the perimeter, how it fragments the opponent), weigh them, and pick the best option. That works well up through Hard. Past Hard, adding more hand-written rules gets diminishing returns — the game stops having a small set of rules that can describe good play.

Elite is what happens when you stop writing rules and let a neural network learn from examples instead. The result is an opponent whose moves are harder to predict and harder to explain, but in a way that feels like playing a person rather than a script.

• Not a cloud service. Elite makes zero network calls while you play.
• Not a subscription. Every difficulty tier is available in the free tier of Jelmata.
• Not random. Same board, same move — Elite is deterministic by design, which makes it easier to study.
• Not scripted. There’s no opening book, no “if-player-does-X-then-Y” tree. Every position is evaluated from scratch.
• Not watching you. Single-player needs no account, sends no telemetry on your moves, and works fully offline.

For developers

The engineering details — how we distilled a full AlphaZero-style teacher down to a 303 KB student, how we exported it to ONNX, and how we got it past Expo’s Android autolinking quirks — live at the Island & Pine studio blog: Shipping a CNN game AI on-device in Expo with ONNX.