Surging Realities | The world model engine

01 · WHY THIS WORKS

World models are infinite factories.

Real-world collection prices an episode in teleop hours and broken hardware. A world model prices it in GPU-seconds: a recent pilot's 40 missing episodes cost $2.32 to generate. That inversion is the entire company.

the data you can't collect

Failure is free in here

The episodes a policy needs most are the ones a real robot can least afford to produce: drops, near-misses, collisions. Inside the world model they cost cents and nobody calls maintenance.

tested before it ships

Every policy runs the gauntlet

The same world that generates training data is an evaluation range. Policies run against the generated failure cases before they touch your hardware, and the readout ships with the weights, whichever way it goes.

sovereign

The weights live in your vault

API labs meter imagination, which is ruinous at training scale and impossible where data can't leave the building. We deliver trained weights on your data, on your cluster if you want. Your simulator is your moat.

02 · THE ROBOT RECEIPT

Cosmos 3 Edge, fine-tuned in ten days.

NVIDIA released Cosmos 3 Edge on July 20 with no public path to fine-tune it for a robot other than their own. Ten days later ours ran on an SO-101, for $342.

Release to a real arm: the road rebuilt first10 days

Campaign cost: spot GPUs, hard ceiling$342

Dreams obey orders: decode to commandswithin 4% of real

Synthetic delivery: ordered by another agent33 eps · $1.32

Run backward, it's a policy: first real graspday five

a dreamed episode: one real photograph plus 240 commanded actions in, sixteen seconds of imagined pick-and-place out.

Play the live demo ↗

03 · THE CREDENTIAL

Frontier recipes, industrialized.

When General Intuition and Kyutai released MIRA, the first open multiplayer world model, we rebuilt it from scratch in the week that followed: codec, model, four-player fine-tune. MIRA Mini.

four players, one world model: MIRA Mini, our one-week reproduction of General Intuition & Kyutai's MIRA (open release, demo only).

Play it live ↗

Codec: ours against the paper's 29.728.6 PSNR

1B world model: at 52% of the paper's budget12.8 gFID

FlashDreams runtime: bit-exact, same GPU2.75×

Apple silicon: parity-gated MLX porta 2021 MacBook

Open release: report · weights · one-click playerCC BY-NC-SA

04 · THE OPTIMIZATION CURVE

Distilled until it runs anywhere.

Local weights are a promise about hardware, so we spend real effort making models smaller and faster. An 18M-parameter policy runs on the factory cell; the curve below is the same discipline applied to our world models, measured:

FlashDreams runtime: bit-exact CUDA-graph fusion, measured9.3 → 25.7 fps

Self-distillation: two forwards replace nine, shipped2-step model

364M student + small decoder: the consumer-hardware unlock3.2× fewer DiT FLOPs

Apple silicon: parity-gated MLX port, running todaya 2021 MacBook

Endgame: WebGPU in the browser: our founding stack, waiting for models to shrink into itno server at all

real output from our first multiplayer local world model, released in March 2026.

Play the live demo ↗

05 · THE ROAD

White glove today, self-serve at the end.

NOW

Pilots, end to end

One task per pilot: world model built from your demonstrations, missing episodes generated, policy trained and tested, weights delivered. First contract signed.running

Policies by default

Every pilot returns a deployed action model, and teams that train their own models take the data and evaluation layer alone. Ten paid pilots delivered end to end is the bar.in progress

THEN

The self-serve platform

Order certified episodes and train policies inside our worlds through the API, without talking to us. The first lab already trains through it today.first users live

06 · THE ASK

Bring us demonstrations. Leave with a policy.

Success is measured one way: your policy, trained on real plus ours, beats real-only on your own eval.

Pilots

A pilot starts with the demonstrations you already have: teleop episodes, video with control logs, telemetry. You get back the generated episodes, the evals, and a trained action model, via API or as weights on your hardware.

Request a pilot ↗

Teams that train in-house

Your model is your moat; keep it. We supply the data and evaluation layer: the edge cases your fleet logs are thin on, and a gauntlet every model update runs before it ships. When a task falls outside your stack, the same pipeline hands back a full policy.

Start the conversation ↗

Prior work

15 players, 1 world model the first playable multiplayer world model (March 2026) ↗ play.alakazam.gg world models in production, in the browser (2026) ↗ worlds.alakazam.gg verified counterfactual data for robot learning (2026) ↗

the world model engine · by the team serving world models in production at play.alakazam.gg