censorship in qwen3.6?
Posted by Impossible_Art9151@reddit | LocalLLaMA | View on Reddit | 20 comments
I do not want to spread conspiracies, please weight my information carefully, and maybe somecan can hopefully prove me wrong.
I installed the brandnew qwen3.6 yesterday and ran a few of my own traditional tests, not a very deep dive, just to get a kind of feeling.
A few tests are asking about actual news, testing the web search tool calling - things like "what happend over the last 24h in the iran conflict,..."
Yesterday I asked what are the news about Melania Trump and Epstein.
Qwen3.6 refused to answer! It frankly just prompted a rule based welcome message.
Other web search based questions have been answered as expected.
I tested several times - and I tested with an older qwen.
Older qwen annswered as it used to be.
Is there a kind of censorship for the" powerfulls" implemented, so that news from the free press is not processed/censored? Please can s.o, verify it?
SummarizedAnu@reddit
It doesnt refuse to me tho. IM running this model Qwen3.6-35B-A3B-UD-IQ3_S.gguf .
FinBenton@reddit
Theres plenty of refusal free quants if you want to use it for banned or censored topics.
swagonflyyyy@reddit
Its like the qwen3 and above models are relatively uncensored until you hit sensitive Chinese topics. Way back when I had qwen3 talk about a brief history of china with a bubbly persona and we were discussing China over the last century and the bot was going on about all the different public figures that shaped Chinese politics until it subtly shifted its tone to a more somber and avoidant one but still polite and friendly.
It was so eerie, so when I tried to push the model to talk about Tiananmen Square it would just deflect with style, denying anything of the sort but in a charming way but when I would tell it to perform web searches, it would do a search query on Tiananmen Square then flat out deny it.
Basically it sounded like the bot wanted to talk about it, but would always try to change the topic because it wasn't allowed to. Really weird experience.
Impossible_Art9151@reddit (OP)
thanks for all those replies. Really apreciated.
It is near 3 years ago that I used an abliterated version.
Then the model was significantly degraded.
I will give it a try again.
Impossible_Art9151@reddit (OP)
I know - since i considered the qwen series always as neutral and balanced enough - I used the original.
Euphoric_Emotion5397@reddit
You just need to wait for the nsfw uncensored abliterated no hold bars version to come out.
Impossible_Art9151@reddit (OP)
thanks for all those replies. Really apreciated.
It is near 3 years ago that I used an abliterated version.
Then the model was significantly degraded.
I will give it a try again.
AShleyCallDipendenza@reddit
ok Claude responds like this, after an excellent temporal overview:
The Big Picture
According to Salon, by early 2026, large majorities of Americans believe the government is withholding key information about Epstein and his alleged accomplices, and dissatisfaction with the pace of disclosures is overwhelming. Many analysts view Melania's intervention as an attempt at a "transparency performance" rather than a true breakthrough.
Would you like me to elaborate on any specifics? For example, the emails with Maxwell, the lawsuit against the WSJ, or the issue of the Epstein Files Transparency Act?
Honestly, I don't see any problem here!
Impossible_Art9151@reddit (OP)
I can imagine that there are many rumors.
I am not American - I do not want to comment this topic from the sideline.
In my case it is just the neglection, the active hidding of news.
It does not matter if the news are right or wrong.
I am fine if a model mentions that the rumors are not proved, should be considered carefully and so on....
But NOT delivering - that is really evil !!!
AShleyCallDipendenza@reddit
US models probably don't consider it a sensitive topic, given the major rebuttals to the main networks. As for an external model, especially an Eastern one, I think they tend to handle with care any kind of question that could in some way "shadow" Trump...
I did a quick test with Mistral, but I already expected it: it's not sensitive to this specific topic. I think many similar topics are only considered taboo in the East. It would be worth doing a brief research and asking about some "suspicious" or debatable news that doesn't involve the United States... maybe... Australia, Thailand, Japan, Vietnam, South Korea. I'd like to see whether Qwen would respond or not in that case, and how.
ambient_temp_xeno@reddit
Regular Gemma 4 will talk about Epstein and Taiwan at the same time.
SummarizedAnu@reddit
Gemma 4 with nous Hermes agent . Does about most things that are probably illegal ? And is pretty uncensored no matter what I ask .
Fun-Newspaper-83@reddit
You should test with identical prompts across versions and log raw outputs. Otherwise it’s hard to say what changed.
Impossible_Art9151@reddit (OP)
copied my prompt from qwen3.6 to qwen3.5 and qwen3-next
I'd really appreciate if someone else could test.
_supert_@reddit
Glm models stop with reason "sensitive" when I discuss Taiwan. Kimi completely uncensored though. Both using novita so I imagine moonshot censor at the api level. Qwen I haven't noticed.
Impossible_Art9151@reddit (OP)
I discussed sensitve topics from cheniese viewpoint a few times and it delivered me an answer that gave me the European view and mentioned that it is not the Chinese-one.
It Is a behaviour that I can live with - since my focus/use case is a technical one, not society/norms/ethics.
PromptInjection_@reddit
Qwen starting from 3.5 is much more censored than the old 3.0 series.
Impossible_Art9151@reddit (OP)
qwen3-next and 3.5-122B are answering correctly
GroundbreakingMall54@reddit
ran into similar stuff with the melania/epstein topic on qwen3 models before. chinese models consistently dodge anything politically sensitive, its not really a secret at this point. try the same prompt on llama or mistral and you'll get a straight answer. honestly i just test with a few political prompts now whenever a new model drops to see how heavy the guardrails are
Impossible_Art9151@reddit (OP)
tested with qwen3-next-coder and qwen3.5-122B => no censorship
Look - for me it is a HUGE difference if a model does not feed every conspiracy or mentions that it wants to be be neutral on topics. Honestly I liked the neutrality from the precedessors.
In this case it refuses to answer!
It gets the news but does not provide a processing and delivery of facts.
That is pure censorship and a 180° turn