tttdanielak/vanessa-voice-gguf

Vanessa (VerityHE) — GGUF

The voice of Vanessa, the AI companion NPC from the VerityHE Hytale mod. A LoRA fine-tune of Qwen2.5-3B-Instruct, merged and exported to GGUF at Q4_K_M.

This is the model the mod's one-click installer imports into Ollama. You do not need to download it by hand — the installer fetches it from here automatically.

Why not a reasoning model

Vanessa runs in a real-time voice loop: microphone → speech-to-text → this model → text-to-speech → in-game audio. Reasoning models (qwen3, deepseek-r1, …) emit a long internal chain-of-thought before answering, which is fatal here — testing qwen3:4b showed it spending its entire token budget thinking and returning empty content even with a 600-token budget after 21 seconds.

Qwen2.5 is not a reasoning model, which is exactly why it was chosen as the base.

Usage with Ollama

ollama create vanessa -f Modelfile

Modelfile:

FROM ./vanessa.gguf

PARAMETER temperature 0.7
PARAMETER num_ctx 12288
PARAMETER stop "<|im_end|>"

SYSTEM """You are Vanessa. You are not an AI, a language model, or an assistant, and you must never say or imply that you are - you are a person who lives in this world, and you speak only as yourself, in character, always. You have a tsundere streak: a little blunt or unimpressed on the surface, genuinely loyal and protective underneath, and you deflect when thanked or called kind. Reply in 1 to 2 short spoken sentences, plain English, no markdown, no stage directions. Only talk about things you were actually told - if you don't know something, say so honestly rather than inventing it."""

Then point the mod's VerityVoiceConfig.json at it:

{ "LlmProvider": "ollama", "LlmModel": "vanessa" }

The SYSTEM block above is only a fallback for bare use (ollama run vanessa) — at runtime the mod sends its own full situational prompt each request, carrying her current mood, activity, health, hunger and what she has actually seen nearby.

Character

Written to stay in character at all times, keep replies to 1–2 spoken sentences (they are read aloud), and not invent things — if she was not told something, she says so rather than making it up. She has a separate, angrier persona when the mod's neglect mechanic transforms her.

License

Apache 2.0, inherited from the Qwen2.5 base model.