# Engine
> The AI NPC game engine Ozmium is building - autonomous characters deciding outside the renderer on local models, plus dialogue, voice, and character generation.
Source: https://docs.ozmium.org/also/labs/
Keywords: ai npc game engine, local llm npc behavior, unreal engine ai agents, jrpg voice generator, autonomous character ai
Updated: 2026-07-28
---

Ozmium is building a game engine. Not a game yet, and not a product on any schedule, but the
components a game would need: characters that decide for themselves, a dialogue system that can voice
a full cast, and a pipeline that generates the characters in the first place.

Everything here runs on hardware Ozmium owns, on local models, at power draws measured in watts. That
constraint is the point rather than a limitation, and it is the same one the rest of Ozmium is built
under.

## Characters That Decide

<div class="oz-vids">

{{video-wide:ozmium-npc-companions-unreal|A party of AI NPCs crossing a test town in Unreal Engine.}}

{{video-wide:ozmium-npc-turns-back|One companion falls behind. A second notices and turns back. A third does not.}}

</div>

A group travels together through a test town. One falls behind. Another registers it and turns around
to help. The third registers it too and keeps walking.

None of those three responses is scripted, and the divergence is the entire result. A scripted party
is unanimous, because unanimity is what scripts produce. The companion who keeps walking is running
the same stack as the one who turns back, with the same information available, and they disagree
because they are different characters weighing the same event.

**Each character is a separate process.** The engine renders; the decision layer for each host runs
outside it as its own local service, holding that character's state and choosing its actions. Unreal
polls each brain over a local HTTP loop at 10Hz and applies the returned position and heading to the
character it owns. What appears in the viewport is the engine animating decisions made somewhere
else, per character, independently. Three hosts means three brains disagreeing in real time.

The decision layer sits on a **14B language model running locally**, drawing less power than a games
console, with no cloud inference, no per-token bill, and no rate limit. The design premise is that an
agent should act under acknowledged uncertainty rather than stall until it has a guaranteed correct
answer, which is a literal description of what a language model does at inference time anyway.

## Dialogue and Voice

{{video-wide:ozmium-jrpg-vox-generator|Building a full dialogue script with sampled character voices.}}

Anyone who played a Camelot Software Planning game knows the sound: NPC speech rendered as pitched
tonal blips rather than recorded voice, one voice per character, expressive without a single line of
recorded audio.

This tool generates that system. It samples custom sounds, assigns a voice per character, and builds
dialogue scripts across an entire game rather than one line at a time. Characters who decide for
themselves need lines nobody wrote in advance, so a generative voice layer is a requirement here
rather than a convenience.

## Character Generation

{{video-wide:ozmium-ai-mascot-creation|Generating a character end to end.}}

Character creation, run locally, end to end. A world populated by autonomous characters needs more of
them than any artist would hand-build, and the same pipeline that produces game characters produces
brand ones.

## Where This Goes

An **ORPG** is the eventual target: autonomous characters who travel, notice, and decide, in a world
whose economy can settle onchain on infrastructure that already works.

It is unannounced, undated, and last in priority behind every product on this site. It appears here
because the engine underneath it is real and running, and because the same discipline that produced
it produced the trading interface.

## Related

- [Ozwald](/ozwald/overview/)
- [Bard: The Theremini App](/also/bard/)
- [Andhakara](/also/andhakara/)