An independent developer has demonstrated Varkos, an experimental AI gaming companion designed to do more than conduct character dialogue. Running in Skyrim, the system can follow conditional instructions, interact with objects, join combat and maintain plans whose next step depends on later events in the game world.
Examples published by the developer include waiting for an arrow signal before retrieving a potion, searching for a named sword, playing hide-and-seek and gathering multiple objects for the player. The planning layer preserves targets, monitors progress and can repair or stop a plan when the game state changes. Its actions are grounded in references supplied by the game rather than relying only on generated conversation.
Varkos is presented as a demon reincarnated as a dog, although the runtime and character are configurable. The starting traits are authored, while a separate process uses shared experiences to alter explicit traits, emotional tendencies and parts of the character's vocabulary. Those changes are versioned and reversible. Personality evolution still uses larger cloud models, but it runs outside the immediate action path.
For responsive play, the project mixes smaller models with conventional language processing, classifiers, rules and game behavior. Its Action Latent Encoder interprets commands, negation, sequence and targets while incorporating a JSON representation of world state. The developer reports that this component takes roughly 2 to 20 milliseconds on an M4 MacBook. A local language model then combines the selected action with personality, emotion and recent history.
The game runs on Windows while audio processing and the companion's main logic run on the developer's MacBook. Speech recognition uses an optimized Qwen3-ASR 1.7B setup that processes rolling audio segments, with a stated target of 40 to 80 milliseconds. Speech generation uses optimized PocketTTS or Qwen-based components. In selected cases, the character can start answering in under 500 milliseconds after the player stops speaking, according to the project account.
The developer says the architecture could run entirely on Windows with about 12GB or more of dedicated GPU memory. Current limitations include drift in longer conversations and a game-specific adaptation step for the action encoder.
Varkos is intended eventually to move between games. When no game is active it enters a speech-only void mode, preserving the character between environments. Parts of the system, and later potentially the full project, are planned for open-source release, but no release timetable was provided.
The approach is deliberately hybrid rather than model-only. Fast local components handle perception and action selection, while slower generative work is kept away from the most time-sensitive loop. That separation is the project's answer to the cost and delay the developer encountered when larger models controlled every decision.
