Qwen3.5

[thaiminhlong]

Hi, I just saw that there's a new model dropped from qwen today and it claims to support up to 201 languages. I'm curious to see how it performs, could you able to do a quality benchmark for this? Qwen3.5.

[hydropix]

Hi, I tested two versions of Qwen 3.5 (9b and 35b), and they give exceptional results for Chinese, but less good results on average for all languages. This is often the case with Chinese models, which have better quality corpora in Asian languages. Strangely enough, it is also very good for Spanish. In short, it is a contrasting model, but a very interesting one.
https://github.com/hydropix/TranslateBooksWithLLMs/wiki/Model-qwen3-5-35b

[thaiminhlong]

Thank you, the result is interesting. Like how the 9B model is better than the 35B model (for Chinese Simplified), I guess I have to try it myself. But anyway, I'm really grateful for this benchmark to exist because I don't really see many benchmarks relating to translating.

[hydropix]

The result for model 9B is most likely a statistical anomaly; this is the first time I have seen such an obviously abnormal result. I would need to test the same translation several times to obtain a more consistent result. In a few months, I will perform a complete update of all models, using a slightly more robust methodology and a higher-capacity LLM judge.