Lllama 4 doesn't output Chinese or Japanese characters

redpnda

Llama4 Maverick fails to output Chinese or Japanese characters under specific query conditions, despite functioning correctly when asked directly in a new chat.

There is a conversation (I hope I don't forget not to delete this):
https://kagi.com/assistant/fea77a6e-9a62-4ce7-b5c7-6138ad87a90c

To replicate, ask Llama 4 Maverick somewhat:

tell me about half slugs
do they really known as half-slugs?
what is chinese term for that?
use chinese characters
what about japanese?
use kanji

The weird thing is that if you ask directly in a new chat, it returns with the right answer, with all characters in place:

give a chinese term for semi-slug

Llama4 Maverick should return responses with the correct Chinese characters or Kanji.

Thibaultmol

@yiwei-1 Can you have a look?

@redpnda I do want to quickly say that the llama 4 models in general aren't that great anymore, depends on your use case obv, but I would personally recommend trying some of the Kagi recommended models and such, they might not have this problem and might actually perform better in genearl

redpnda

Thibaultmol I'm a novice in AI assistance. I've tried several models and noticed that Llama4 is quite fast in comparison to DeepSeek V3, for example, which attracted me to it.

Hanbyeol

I think this is more of a limitation of the Llama model itself rather than a Kagi issue.
I tried via OpenRouter but it also didn't show letters:

On the other hand, Gemini 2.5 Flash Lite (an even lighter model than 2.5 Flash) shows letters.

(To be precise: Gemini Flash Lite is writing Simplified and Traditional characters on the same line, which is the WRONG answer, but at least the characters are displayed.)

So it seems Meta hasn't properly trained its model for Chinese and Japanese, given that they don't mention those languages either.
https://www.llama.com/docs/model-cards-and-prompt-formats/llama4/

redpnda

I'm not sure where the issue is. The direct question works more or less. You can try it like this:

I understand that the problem may be in the model itself. Anyway, it's good to know that some models are superior in certain fields.

yiwei-1

As @Hanbyeol has shown, this is more of a model limitation with Llama 4. You can try other strong multilingual models available, such as Qwen3/Kimi K2/DeepSeek v3/etc.