Choosing AI Providers (Local & Cloud)
Configuration guide and benchmark comparisons across local offline LLMs and cloud API translation providers.
Preview Content
Please note that the details on this page are pre-content drafts. More detailed technical specifications, architectural breakdowns, and updated guides will be added soon.
1. Configured AI Providers
XianScan features a universal OpenAI-compatible LLM client runtime with process-wide queue concurrency control, auto-retry, and reasoning tag suppression.
| Provider | Default Base URL | Recommended Active Models | Best For |
|---|---|---|---|
| DeepSeek (Default) | https://api.deepseek.com | deepseek-v4-flash, deepseek-v4-pro, deepseek-chat | Asian idioms, Xianxia / Murim terms, fast 1-2s response |
| Google AI Studio | https://generativelanguage.googleapis.com/v1beta/openai/ | gemini-3.7-flash, gemini-2.5-flash, gemini-2.5-pro | Massive context windows, high batch throughput |
| Groq (Ultra-Fast) | https://api.groq.com/openai/v1 | llama-3.3-70b-versatile, llama-3.1-8b-instant | Sub-second real-time generation speed |
| Ollama (Local) | http://localhost:11434/v1 | qwen2.5:14b, qwen2.5:7b, qwen2.5:32b, deepseek-r1:14b | 100% Offline, private local GPU translation |
| LM Studio (Local) | http://localhost:1234/v1 | local-model, qwen2.5-14b-instruct-gguf | Offline desktop LLM interface with GGUF quantization |
| OpenAI | https://api.openai.com/v1 | gpt-4o-mini, gpt-4o, o3-mini | Consistent syntax and standardized English prose |
| OpenRouter | https://openrouter.ai/api/v1 | anthropic/claude-3.7-sonnet, deepseek/deepseek-v4-flash | Universal multi-model routing and fallback pools |
| Custom | http://localhost:8000/v1 | User defined (e.g. vLLM / LocalAI / SGLang) | Self-hosted high-throughput GPU clusters |
2. Setting Up Local Ollama
- Install Ollama on your system.
- Pull a recommended multilingual comic translation model:
bash
# 14B parameter model (recommended for 12GB+ GPUs):
ollama pull qwen2.5:14b
# 7B parameter model (recommended for 6GB - 8GB GPUs):
ollama pull qwen2.5:7b
# Reasoning-distilled model for complex narrative prose:
ollama pull deepseek-r1:14b- In XianScan, go to Settings -> AI Translation Providers -> Ollama (Local).
- Set the active model name to your pulled model (e.g.
qwen2.5:14b). - Click Test Connection.
3. Context-Aware Dialogue Memory & Tracking
XianScan employs a sliding-window cross-page dialogue context tracker (DialogueContextWindow) during translation. The tracker preserves up to 5 previous pages of speaker identity, pronouns, and topic context to maintain character voice consistency across page breaks.