One save.
Two futures.

The world is a file.

In a neural game there is no level geometry, no entity list, no scene graph. Every frame is sampled by a video diffusion model, and the only place the world exists is the model's running state: a few gigabytes of attention cache, a sliding window of latent frames, and the exact position of the random-number generator.

.wsave serializes all of it. Load one back and the world resumes bit-for-bit — the same clouds, the same grass, the same next frame. Neural games never had a save button. Now they do, and it turns out to be a measuring instrument.

An honest footnote: naïvely quantizing the cache to int8 shrinks the file to 1.9 GB and ruins the world (13.5 dB) — the state is far more fragile than the weights. A 4.8 MB re-prime save matches that fidelity at 1/400th of the size. The gap between "exact" and "re-primed" is a direct measurement of what the cache remembers beyond the pixels.

Your opponent is a probability distribution.

Freeze a world. Resume it twice with the same actions but different noise: the copies drift apart with a half-life of 109 frames — 4.4 seconds. Nothing in the game scripts this; it is the raw stochastic physics of the model.

Now do the opposite — same noise, different action — and the worlds separate 3.55× faster. That single ratio is the entire premise of the game: your hands are stronger than chance. But chance never stops.

Kindred instrument: Hansen & Wang detect world-model hallucination live with label-free predictors and dare you to break their world. Butterfly is the mirror image — chaos is always breaking ours, and the game is to hold it. Their flow-instability meter is on our roadmap as a live "dream stability" gauge.

The mission: your world is about to be re-rolled.

Butterfly's default run has one explicit goal. You get 10 moves to walk — that walk becomes the original timeline, the world you must defend. Then THE TEAR: fate re-rolls, and chaos starts pulling the dream away from what it was.

Hold it for 10 moves. Each move where you push the world further than chaos itself would (drift > 1.25× the measured baseline) is a strike — three strikes and the dream tears. Survive, and you're graded S/A/B/C by how much more world you held than doing nothing. The frame's border burns green → red with your drift, live. Mimicry scores zero; you win by steering against the drift.

Play against the physics, not a script.

Every mechanic below exists only because the engine is a neural network. A classical engine cannot offer any of these; a neural one cannot avoid them.

Runs on one GPU. Honestly.

Butterfly drives Matrix-Game-2.0 (1.8B, MIT) as its first world model host — 29.5 fps at 352×640 on a single H200, ≥24 GB VRAM. No hosted demo yet; you bring the GPU, we bring the save button.

# 1 · code + world model source
git clone https://github.com/lixuan27/butterfly && cd butterfly
bash scripts/get_matrix_game.sh          # clones MG2 + downloads weights (~28 GB)

# 2 · serve the dream (needs 1× GPU, ≥24 GB)
bash scripts/serve.sh                    # http://localhost:7860

# 3 · play
WASD walk · IJKL look · ⚓ anchor · ⚔ duel chaos · 🦋 one flap · ⇩ relay a dream