Kagi Translate

Translation quality and benchmarks

  • Some users are impressed, citing better handling of specific ambiguities (e.g., “Orgel” vs internal organs, Pig Latin, Romanized Arabic) and certain timing expressions where Google fails.
  • Others report poor or stilted output in Afrikaans, Turkish and Cantonese, with Google or DeepL doing better and more idiomatic translations.
  • Several comments highlight refusal/censorship instead of literal translation for profanity or sensitive content; this is seen as a significant negative, especially when not clearly indicated.
  • Multiple people question the “better than Google/DeepL” claim and ask for rigorous benchmarks (e.g., BLEU scores, dataset sizes). Past Kagi benchmarks are criticized as too small.

Language support and copying concerns

  • Users diff the language lists and find minor mismatches with Google (e.g., Crimean Tatar, Santali variants, Inuktut, Tshiluba).
  • The near-identical lists lead to speculation that Kagi largely copied Google’s language menu.

Implementation details

  • Kagi states it uses a combination of LLMs and selects the best output.
  • Commenters speculate this likely means routing between existing large models (Qwen, Llama, OpenAI, Gemini, Claude) rather than fine-tuning their own.

UX, features, and limitations

  • Lacks “translate as you type” and feels slower than DeepL; Kagi says live translation will come, likely as a paid feature.
  • Webpage translation only handles initial HTML, not content added later via JavaScript; iframes and x-frame-options often block page mode.
  • Users want: explicit context/tone controls (formal/casual, gender, T-V distinction), pronunciation hints (e.g., for Japanese), multiple variants (like Bing), an API, Android app integration, and better single-word handling.

Access, captchas, and bots

  • Cloudflare Turnstile frequently misclassifies humans, blocking both Translate and the feedback forum; this frustrates paying users.
  • Some argue Cloudflare is a necessary cost-control measure for a free, compute-heavy service, but others note Google avoids similar friction.

Kagi’s broader product strategy

  • Several users appreciate Kagi search but feel Translate, Maps, and the browser are “quarter-baked” and dilute focus.
  • Others accept this as a resource tradeoff in a user-funded, long-term plan, but suggest clearer “beta” labelling and setting expectations, especially for weaker offerings like Maps.