Skip to main content
Choose output settings in the processing selector or Settings → Translation. Koharu captures them when a run starts; edits affect later runs. Provider connections and credentials are read for each request.

​
Language and instructions

Set the target language and concise instructions for names, terminology, tone, pronouns, formatting, or text that should remain untranslated. Instructions are saved with Koharu’s settings and apply to every project; there is no per-project field. Test them on several pages and correct OCR first. LFM 2.5 and Ministral 3 support a limited set of target languages, and a run fails when their target is unsupported. Other local models accept every target language Koharu offers.

​
Local quantization

GGUF models may offer multiple quantizations. Smaller files need less download space and memory, usually at a quality cost. Larger files can be slower or exceed available memory. Hosted providers control their own serving precision.

​
Generation controls

Start with provider defaults and change one variable at a time. High randomness tends to make names and terminology less consistent.
Reasoning can increase latency and token use. Reasoning and vision are dropped silently for a model that does not support them; other unsupported options may be ignored or rejected by a provider or model.

​
Text and vision inputs

Text-only requests send semantic source text. With Vision, a capable model also receives the source page image. A custom OpenAI-compatible server is assumed to accept images, so turn Vision off if it does not support image messages. Generated translations remain editable project data. Review the wording and page layout before export.