01The Bet

A specialist model
that runs inside
the live-ops loop.

Stealth until summer 2026

Talk to us → The model
02What we do

We've trained a model
on what actually happened.

SuperTuned is a base model and a growing library of specialist adapters, trained against the multi-dimensional preference real operators run their games on — retention, monetization, fairness, feel. Every adapter is post-trained inside the environment of the studio whose work it captures. Studios keep their data. We ship the adapter. The library compounds as the industry does.

Trained on what shipped, what stuck, and what converted.
03Why it has to be us

Four prerequisites.
One team that has all four.

Data

200 million players' worth of live behavioral signal.

Not licensed, not scraped, not borrowed. Studios keep custody — we train adapters inside their environment.

No telemetry leaves the studio.

Domain

Twelve years operating live economies. Sixteen years shipping games.

Co-founders for twenty-four years. Every company since 2002.

Method

A post-training stack run with our training partner.

A shared open-weights base, plus a library of per-studio adapters tuned against operator-elicited preference.

Distribution

First commercial deployment, summer 2026.

Stealth until launch.

Most teams claiming to build AI for games have one of these. We have all four.

04The model

A base model.
A library of adapters.
One layer for every studio.

An open-weights base in the 8B–30B parameter range, with per-studio low-rank adapters that hot-swap at serve time. Each adapter is trained inside the studio's environment against the multi-dimensional preference that operator runs on — retention × monetization × fairness × feel. Studios keep the data. We keep the library. Cross-studio compounding lives in the library, not the base.

SuperTuned stack — Studios on top, SuperTuned in the middle, Understudy underneath A three-layer stack diagram. Top layer is studios running their own adapters. Middle layer is SuperTuned, which holds the per-studio adapter library on a shared open-weights base model. Bottom layer is Understudy, which provides the infrastructure: supply routing, post-training pipeline, and serving stack. STUDIOS Each studio runs its own adapter. Telemetry stays inside the studio. The adapter serves inline. GAME A GAME B GAME C ··· ↑ ADAPTER SERVES ↓ TRAINING SIGNAL (NEVER LEAVES) SUPERTUNED A base model. A library of adapters. The library is the compounding asset. ADAPTER A LoRA · STUDIO-TRAINED ADAPTER B LoRA · STUDIO-TRAINED ADAPTER C LoRA · STUDIO-TRAINED ··· SHARED OPEN-WEIGHTS BASE 8B–30B PARAMETERS UNDERSTUDY Infrastructure underneath. Supply routing. Post-training pipeline. Serving stack. SUPPLY ROUTING POST-TRAINING SERVING STACK
200M+
Players in the training signal
16 yrs
Shipping games
12 yrs
Operating live economies
24 yrs
Founding partnership
4 / 4
Prerequisites met
Summer '26
First deployment

One base. Many adapters. One library.

05On the name

Tuned is fitting to the past.
Super-tuned is fitting to the preference behind it.

Most specialist models are trained on (input, output) pairs from historical work. They curve-fit to the surface — the artifacts a team shipped, not the judgment that produced them. SuperTuned is post-trained against an operator-elicited preference signal that captures the multi-dimensional shape real operators run live products on — retention × monetization × fairness × feel. The preference is the asset. The adapter is the artifact.

Tuned

Fitted to historical designs. Strong inside the portfolio. Weak outside it.

Super-tuned

Fitted to the underlying preference. Generalizes across genres, formats, and stacks.

Without preference-tuning, every model is a curve-fit. With it, we get a generalist for the specialist work.

06What's next

A tuning layer
for the games industry.
Quietly.

One adapter goes live in August.
The next ones get easier.
If you're running live signal of your own, we should talk.

Talk to us →

hello@supertuned.ai · stealth until summer 2026