Choosing a Model
Pick the right model for translation quality, speed, and cost across providers.
The model you choose affects translation quality, speed, and cost. Here is what to consider for each provider type.
Local Models (LM Studio / Ollama)
Good starting points
Dedicated translation models such as HY-MT1.5-7B or NLLB-200 variants
translate well out of the box. General instruct models like llama3.2 or
qwen2.5 (7B class) are also solid all-rounders and handle Karpa's tone
instructions nicely.
| Factor | Guidance |
|---|---|
| Quality | Larger and less-quantized models translate better. |
| Speed | Smaller, more heavily quantized models are faster. |
| Memory | Match model size to your available RAM/VRAM. |
| Languages | Some models are tuned for specific language pairs; check the model card. |
Quantization explained
Quantization compresses a model to use less memory, with a small quality trade-off.
| Level | Trade-off |
|---|---|
Q4_K_M | Sweet spot: good quality, low memory. |
Q5_K_M | Slightly better quality, a bit larger. |
Q8_0 | Near full precision, largest footprint. |
Cloud Models
| Provider | Default Model | Good For |
|---|---|---|
| OpenAI | gpt-4o-mini | Fast, affordable, great quality |
| Anthropic | claude-sonnet-4-20250514 | Nuanced translations, longer context |
| Google Gemini | gemini-2.0-flash | Fast, generous free tier |
| OpenRouter | Your pick from 200+ models | Variety and competitive pricing |
Cloud defaults can be changed in Settings → Connection at any time.
Trying a new model
- Local: download and load the model in LM Studio or pull it with Ollama, then select it in the model picker.
- Cloud: just select it from the model picker in Settings.
- Run a test translation on text you know well.
If translations break after a model swap, the model identifier probably does not match what your provider exposes. Double-check for typos, or re-fetch the model list from Settings → Connection.
More fixes live in Troubleshooting.