LLM & Translation Services Overview & Comparison Matrix
In DuRT 3.0, Large Language Models (LLM) power Subtitle Studio interactive restructuring, colloquial refinement, meeting minutes extraction, and contextual translation, while acting as the core intelligence in the DAG workflow pipeline.
Combined with macOS 15 native offline neural translation, DuRT delivers a dual-track experience: ultra-fast on-device live bilingual subtitles and deep contextual cloud translation.
4 Model Camps Matrix Comparison
DuRT uses a pure client-side BYOK (Bring Your Own Key) architecture. Requests are sent directly from your Mac to model providers with zero proxy markup and zero data interception.
| Category | Providers & Engines | Models | Max Context | Reasoning / Thinking | Vision | Network Req. | Best For |
|---|---|---|---|---|---|---|---|
| 🍎 Native Offline | Apple Translation | System Neural Translator | Sentence stream | ❌ | ❌ | 100% Offline | Live internal audio dual subtitles, offline translation |
| 👑 Global Commercial | OpenAI DeepSeek Claude Gemini xAI Grok | GPT-4o / GPT-4.1 DeepSeek-V3 / R1 Claude-3.7-Sonnet Gemini-2.5-Pro Grok-3 | 64k ~ 128k+ | ✅ DeepSeek-R1 ✅ Claude-3.7 | ✅ GPT-4o ✅ Gemini ✅ Claude | Global access | Deep interview logic extraction, high-nuance translation |
| 🇨🇳 Domestic Compliant | Qwen / Zhipu / Kimi / Hunyuan / ERNIE | Qwen-Max / Qwen-Long GLM-4-Flash / Long Moonshot-v1-128k Hunyuan-Turbo | 8k ~ 128k+ (Kimi ultra-long) | ✅ Reasoning options | ✅ Supported | Direct domestic access | Chinese idioms, long audio meeting minutes |
| 💻 Local Private | Ollama Local Models | Llama 3.3, Qwen 2.5, DeepSeek-R1 Local | Limited by VRAM | ✅ Depends on model | ✅ Vision models | 100% Offline LAN | Confidential boardroom meetings, zero data upload |
| ⚡️ Fast Inference & Routers | SiliconFlow Groq OpenRouter | DeepSeek-V3 / R1 API Llama-3.3-70B (LPU) Global Model Router | 64k ~ 128k | ✅ Supported | ✅ Supported | Fast direct access | Parallel workflow batching, high-throughput pipelines |