First token seems bugged out
The first token is not the initial think "" tag and starts spitting out random words before responding after "closing" the thinking tag (which doesn't appear as the first token), after second or third prompt it seems to go back to normal with the thinking tags. Used recommended Qwen3 settings (temperature 0.6).
In LM Studio Failed to parse Jinja template: Parser Error: Expected closing statement token. OpenSquareBracket !== CloseStatement.
In llama cli: common_chat_templates_init: failed to parse chat template (defaulting to chatml): Expected value expression at row 18, column 30: {%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %} {%- for message in messages[::-1] %} ^ {%- set index = (messages|length - 1) - loop.index0 %}
And it speaks Chinese for some reason.
Ah right right .. I didn't apply the template fix from the other models here :')
tried Q6_K_L, Q5_k_L, Q5_K_M all have the same issues
it speaks Chinese a lot, especially in the very first answer,
not very alliterated (if at all)
they all like to add \u200B char randomly in the code, very annoying. it's invisible for the eye...
Didn't encounter that problem with 30B version
template is incorrect (not a big deal, one can use from qwen2.5 or from any working one, just copy paste)
do not follow the previous corrections... so instead of fixing existent ones they create new code with new issues
not usable
Someone said after 1 or 2 outputs, it starts behaving normally, and i can confirm that. But you have to replace Jinja template to a correct one, i did it in LM Studio.