alakazam · engine
The world model engine

We build the worlds robots learn in.

Custom action models for any task. A world model built from your demonstrations generates the failure cases you could never stage, then trains and tests the policy. Your data, your weights, via API or on your hardware.

Request a pilot ↗ Explore →
01 · HOW IT WORKS

From a handful of demonstrations to a policy that ships.

Robot learning is bottlenecked on one thing: the data a real robot cannot afford to collect. The dropped part, the near-miss, the snag. We build a world model of your task and make that data cheap.

Bring demonstrations. Teleop episodes, video with control logs, whatever your robot already records. A recent pilot started from 174 hand-collected episodes of one task.

We build the world model. A model of your task that answers both questions: what happens if the robot does this, and what should it do to make that happen.

It generates what the dataset is missing. New positions, new layouts, the failure cases nobody stages on real hardware. For that same pilot, the 40+ missing episodes cost $2.32 to generate.

The policy trains and gets tested inside the world before it touches your robot. You get the episodes, the evals, and the weights: via API or running on your own hardware.

02 · THE FIELD RIGHT NOW

Ten pilots in fourteen days.

First contract signed, two MOUs. A sample of what is running today, described without names because their data is theirs:

contract signedAn autonomous factory

An 18M-parameter assembly policy trained inside a world model of their demonstrator. Small enough to run on the cell, tested on the tolerance failures nobody stages on a real line.

live a/bA university robotics lab

A real two-armed fruit-picking robot, our generated episodes against their baseline data, measured on the robot itself. MOU signed for deliveries beyond the A/B. The readout ships whichever way it goes.

mou signedA cable-inspection drone company

Autonomizing electric-line inspection and repair. The aircraft carries no camera, so the world model builds on drone state and RF sensing, and the approaches nobody flies near a live line get generated, not flown.

in productionA spiking-neural-network lab

Training policies inside our world models through the public API. The gym, in production, today.

scopingA campus self-driving team

Dataset augmentation plus a fine-tune for a self-driving cart at a major university. A published result matters more to us here than the revenue.

your robotOne task, priced in days

A pilot starts with the demonstrations you already have and ends with episodes, evals, and weights. Request a pilot ↗

03 · WHY THIS WORKS

The world model is the factory.

Real-world collection prices an episode in teleop hours and broken hardware. A world model prices it in GPU-seconds. 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.

“The failure cases your robot most needs to learn from are the ones you can least afford to stage.”
04 · THE CREDENTIAL

Frontier recipes, industrialized.

Training and evaluating world models is a rare skill, and we prove ours in public, on the clock. Two receipts:

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 arm, for $342. The road had to be rebuilt before it could be driven.

MIRA Mini, reproduced in a week. When General Intuition and Kyutai released MIRA, the first open multiplayer world model, we reproduced it from scratch in the week that followed: codec, model, four-player fine-tune. Then our runtime work made it 2.75× faster on the same GPU and a parity-gated MLX port put it on a four-year-old MacBook.

Credits: MIRA's architecture, recipe and dataset are General Intuition's and Kyutai's, released openly, in collaboration with Epic Games. The dataset used there is licensed CC BY-NC-SA; that model is for demonstration purposes only.
Read & download
MIRA Mini technical report ↗ Play it live an L40S, ~19 fps ↗ Weights CC BY-NC-SA ↗ The one-click player ↗
05 · 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 ↗
06 · 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

NEXT

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

07 · 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