Does THINKING MODE significantly improve translation?
Posted by Sostrene_Blue@reddit | LocalLLaMA | View on Reddit | 15 comments
Between a solid model from Qwen or Gemma 4, when translating a text, does "thinking mode" significantly boost the quality of the translation, or is the difference negligible?
aboutthednm@reddit
Gemma's English to German and German to English is "good enough", but I wouldn't want to use it in any sort of professional setting. Thinking about it doesn't improve results imho. Sometimes it beautifully translates a long and complex sentence into German, only to be missing a verb or adjective in (to a native speaker) the most obvious place, ruining what would have been a flawless translation. I don't know why it drops random words from sentences that plainly ought to be included, but hey, it's alright. Expecting 100% perfection from LLMs is a recipe for a decline in mental health.
Django_McFly@reddit
Honestly I don't think thinking mode is better for anything that you don't want like a 1-person round robin discussion happening on. Maybe it would be better if you had some context document for how to handle various words or like something that needed more than the raw ability of the model to translate.
taking_bullet@reddit
It's actually otherwise. I use LLMs for translating text from English to Polish and enabled thinking only makes the whole process longer.
Otherwise_Economy576@reddit
language pair matters more than people are saying here. going TO english, thinking adds little because the base models are already good at producing fluent english. the gains from deliberation are minimal compared to the latency cost. going FROM english to a lower-resource target language is where thinking actually shows up. the model uses the deliberation budget to reason about morphology, register, and idiom choices that it would otherwise default to literal substitution on. for english<->european-language pairs both directions are mostly fine without thinking. for english to japanese/korean/arabic specifically thinking has been worth it in my own runs.
t00052e@reddit
I had tried to use Qwen3.6-27B and Qwen3.5-122B-A10B, for translating from Japanese into Chinese. In both models, thinking improves translation quality a lot, although there are still some glitches. However, probably due to limitation in the models themselves, they have difficulty with some context, particularly if the subject or object is omitted.
I have also tried Gemma-4-31b, which has better understanding of Japanese grammar, but have difficulty writing Chinese fluently in some cases, even with thinking is on. Again, I think this is because this model has limited training in Chinese.
I am thinking of using Gemma-4 to do a first round of translation and then using Qwen to repair the translation if the Chinese is broken, when I have the time to do it later.
Temporary-Mix8022@reddit
I have a custom harness and I've generally done:
The second pass has been genuinely useful at picking up some anomalous results. I only speak 2 languages well enough to debug manually, but on Gemma 4 these were real issues.
I have been thinking of adding a pass 1 part 2, where it is literally run again. If there is any disagreement.. it's run again or flagged. Which should average out any weight noise, or flag where it might be less precise.
I then run through a large cloud model, then through a human.. most of this is just trying to cut out the 90% of work before it hits a human or expensive model
Chromix_@reddit
Yes, that is the way, and you can even benefit from partial KV cache reuse on the verification pass.
Temporary-Mix8022@reddit
I have to be honest.. I use Temp=1. I think I saw Google recommend that for all usage.. but unsure really.
Also yep, if you layer the prompts up in the right way, llama.cpp will give you pretty much a 100% cache hit.
Middle_Bullfrog_6173@reddit
My experience is that if you are translating between two non-English languages it tends to hurt. If you are translating to English it tends to improve the output.
If you are translating from English then it depends... At least Gemma's output does improve for complex topics, since it drafts and compares different terms. For other models I've found the opposite, like gpt-oss does better without thinking.
MindPsychological140@reddit
Thinking pays off for idioms, jargon, and long-form consistency. For straight prose it's mostly latency overhead. A dedicated translation tune (Qwen3-Translation, Tower) usually beats generalist + thinking for typical text.
Sostrene_Blue@reddit (OP)
I didn't know that these kind of models exist.
What does it involve in concrete terms?
MindPsychological140@reddit
Concrete: fine-tuned LLMs trained on parallel translation data. Same tooling as any other LLM (llama.cpp, Ollama, vLLM).
The active dedicated line is **Tower-Plus** from Unbabel:
- **Tower-Plus-2B / 9B / 72B**
- Tower-Plus-9B is built on Gemma 2 9B, so memory footprint is what you'd expect from any 9B model
- Supports 22+ languages including EN, DE, FR, ES (+LATAM), PT (PT/BR), IT, NL, RU, ZH (simp/trad), KO, JA, PL, SV, HU, CS, UK, RO, DA, NO, FI, HI, IS
For Chinese pairs specifically there isn't a dedicated officially-released Qwen line in the same way — base Qwen2.5 with an explicit translation prompt works well enough that I'd start there. Community finetunes exist but quality varies.
HuggingFace: Unbabel/Tower-Plus-2B / -9B / -72B
Plastic-Stress-6468@reddit
I find that thinking only improves translation quality when you know what you want. Specifically when you have style or process constraints that you need it to follow.
If you genuinely need translation because you either can't understand the source or the target language of the text you are working with then thinking is usually not worth it. You can't give it specific instructions on pitfalls to avoid anyways, you might as well turning thinking off and back and forth with the model to ask for clarification or changes directly.
LagOps91@reddit
translation should be a relatively easy task and no thinking should be required. be sure that you don't have any repetition penalty set and use low temperature.
UnWiseSageVibe@reddit
I have been using Gemma 4 for translation processing on some personal projects and I found that having it off is better. It wastes a lot of context thinking about it and also ends up overthinking it.