Voxta Large Language Models
Text generation providers — the brain behind every character reply, across cloud and self-hosted options.
The LLM is the AI's brain. It handles three jobs in Voxta:
- Reply — the conversation text the character speaks.
- Action Inference — picking which action to fire.
- Summarization — compressing old chat history into manageable summaries.
You can route all three through one service or split them across different ones (e.g. a high-quality model for Reply, a cheap fast model for Action Inference and Summarization).
The Voxta UI groups LLMs into three hosting tiers — these docs mirror that.
Cloud-Based
External companies. Need an API key. Pay per use.
OpenAI
GPT family.
Anthropic
Claude family.
OpenRouter
Gateway to dozens of models behind one key.
Gemini via OpenAI compatibility.
xAI (Grok)
Grok models via OpenAI compatibility.
NovelAI
Subscription LLM tuned for storytelling.
Mistral AI
Mistral's hosted models.
Voxta Cloud
Voxta's own hosted LLM.
Self-Hosted: Zero-Setup
Voxta installs the runtime and Python dependencies automatically on first use. You only provide a model.
llama.cpp
GGUF models, CPU + GPU.
LlamaSharp
In-process .NET llama.cpp binding.
ExLlamaV3
EXL3 quantized models on NVIDIA GPUs. Fastest local option.
ExLlamaV2
Being removed after 1.11. Use ExLlamaV3.
Self-Hosted: Requires External Software
You install and run the upstream LLM server separately. Voxta connects to its API.
KoboldAI / KoboldCpp
Single-executable local LLM runtime.
Text Generation Web UI
oobabooga's local LLM front-end.
Claude Code
Your local Claude Code CLI and its subscription.
Codex
Your local Codex CLI and its ChatGPT login. Experimental.
OpenAI-compatible
Any service that exposes the OpenAI chat-completions API.