A pilot starts with recordings. Replays, video with inputs, telemetry — a game, a sim, a fleet. You get back a playable world model of your environment, evals against the real thing, and the weights.
Request a pilot ↗a dark pond, illustrated in halftone — every ripple surfaces a world. Orange rings the ones that are playing.
Any game, any recorded reality, delivered as a living, playable world model: your data, your weights, running where you want it.
MIRA, the first open multiplayer world model, was released July 6 with code and data by General Intuition and Kyutai. MIRA Mini is our from-scratch reproduction, built in the five days since: the codec, the model, and the four-player fine-tune, served in a browser or downloadable as locally runnable weights.
What that took: five days on rented spot GPUs.
Then we made it fast. A bit-exact port to our FlashDreams runtime runs 2.75× faster on the same GPU. Our own self-distillation campaign is collapsing nine diffusion steps into two, a 330M student and a compact decoder are in training, and a bit-exact MLX port already runs the model on a four-year-old MacBook. A world you own should run where you want it.
For most jobs the real game engine is cheaper and exact — we will tell you when it is. A trained world model earns its keep on the things an engine cannot do:
Trained on real matches, it absorbed how people actually play. The opponents in our demo were never scripted — they are the training data, alive. An engine cannot answer "what would a player do here?"; the model is that answer.
An engine replays a log deterministically. The model takes any real moment and plays it forward differently: alternate futures with realistic opponents in each. Counterfactual replay has no engine analog.
N rollouts from one state give a probability map over futures: win odds by choice, risk by position. Engines compute outcomes; the model estimates distributions.
You cannot backprop through a game engine. Gradients of futures with respect to actions mean agents train inside the model, GPU-batched. That is the property the embodied-AI market pays for.
Engine states must be reachable. The model needs pixels: start play from a screenshot, an edited frame, a composited scene. And it survives its creator: video plus inputs keeps a world playable after the servers are gone.
“The planet in the glass was never the terrain — it’s the population.”
Ours condition on every participant at once: teammates, opponents, a whole fleet on one floor. Trained on your recordings, the model carries how your people actually behave. That's the part of your world no engine ever had.
The reproduction took days on rented spot GPUs, and the runtime work makes it cheap to serve. The pipeline is amortized; your world rides it. Proven, not promised.
API labs meter imagination, which is ruinous at RL-training scale and impossible where data can't leave the building. We deliver the trained weights on your data, on your cluster if you want. Your simulator is your moat.
Each one survived the filter "why not use the real system?": these buyers either don't have the engine, or need what it can't do.
The agent-training market hand-builds environments at engineer salaries. We turn recordings of play into GPU-batchable, uniform-API environments grounded in how humans actually behave.
Game-agent and QA companies, and studio central tech that already trains replay bots by hand. We are the layer underneath their agents.
Marketing moments now ("play the anniversary, remixed"), and a second life for the replay archive: license your world to the AI-training market, rev-share on data you already own.
Fleets and human-robot floors are multi-agent, which is our native shape. Base models partially serve this lane; the gap is the ops, and delivered weights are the product where data can't leave the building.
Serving is priced in GPU-milliseconds per frame, and the multipliers compose. Measured at the top, in training below:
Honest evals, the war-story writeup, a playable demo; weights follow the rights conversation. The credential.shipping now
A beloved multiplayer game, trained with its rights holder as the first commercially licensable open multiplayer world model. Not a reproduction: an artifact nobody else has.next
Agents trained inside the world model — the application the MIRA paper names first and nobody ships.brief written
Success is measured one way: your policy, trained on real plus ours, beats real-only on your own eval.
A pilot starts with recordings. Replays, video with inputs, telemetry — a game, a sim, a fleet. You get back a playable world model of your environment, evals against the real thing, and the weights.
Request a pilot ↗You own a world people loved. Its archive of play is a model waiting to be trained, licensed on your terms, with the weights under your control. We bring the pipeline. You keep the rights.
Start the conversation ↗the world model foundry · by the team serving world models in production at play.alakazam.gg