Standard translation is unusable for Hungarian

buherator

I've been using Kagi Translate for some time between Hungarian and English and results were really good. Recently I noticed that there is an option to switch between Standard/Best translation while translation quality turned unusable with the default Standard setting.

I just tested with single word translations of dictionary words that should be a trivial task, but Kagi gives nonsense results (e.g. "drill" ([HU]fúró) instead of "anvil" ([HU]üllő)).

When switching to Best the result is correct.

This behavior makes a very bad first impression of Kagi for new users.

I expect Standard translation to be useful at least for trivial tasks, but it'd probably better if Translate just gave good results without additional dials.

Gredharm

Does anything translate Hungarian well though? I've found Kagi better than Deepl and GPT5, and far better than Google Translate. But I've just accepted that Hungarian is beyond current technology, or my own mental capacity.

buherator

@Gredharm I found Kagi very good for Hungarian. For example, I used it to double check a thing I wrote and it even recognized metaphors and offered alternatives I could use depending on my intent! (this may have been from English though). I also used it to translate my long and highly technical posts from Hungarian and the result was beyond my expectation. The "Best" option also works well today, but having Standard as the default can be a total deal breaker for the average user (and pretty annoying for me).

Another info for the report: accuracy seems worse with single-word translations, I guess the fast LLM on the backend can infer better if context is given, but goes off-rails without that.

Gredharm

buherator I wonder if that's the case for many other languages. I assume they use different LLMs for Standard and Best quality translations, and it would make sense that a smaller, faster, cheaper model wouldn't be very good with less common languages.

As such a solution may be to automatically select Best for languages not properly/well supported by the Standard model, and put a small note indicating that.

Thibaultmol

For the 'it being slow for short queries like words' it's because in that case it's doing other stuff (dictionary stuff) I think

buherator

Thibaultmol By "performance" I didn't mean speed but accuracy. I edited my post for clarity.

(As it often is with LLMs, if they simply used an orthodox method like a dictionary lookup for single words it would be faster, cheaper and more accurate.)