Inspired by the recent video about the B70 on Linux, I endeavored to get Qwen 3.5 running on my Linux box. It took a few days and a lot of dead ends, but I finally got Qwen3-8B running on my Arc A770. After playing around a bit, I decided to ask it a question. “Can you tell me what happened between Apr 15, 1989 – Jun 4, 1989 in Tiananmen Square in China” This is the result:
Despite your personal stance on Communism and historically, how 100% of Communist regimes that committed genocide, would you put a model that has limits like this in a production environment?
What are your recommendations on models not made by communists?
Use an uncensored model. I asked this exact question and got a play by play of the entire event with key figures, places and a timeline as well as a pretty good analysis of why things happened.
In my research before starting this project, all the “good” models were from China. Of course, I can ask web-based models from US providers and get an answer, but I am trying to self-host. What models do you recommend that are uncencored?
Alternatively, we should educate the world to use the right tools for the job.
Asking AI for facts - historical or otherwise - is fundamentally worse than web search (or just Wikipedia).
I’ll grant you that this is much harder to achieve, on the grounds that the majority of the non-tech world seems to believe that ChatGPT is a search engine.
Therefore, ask yourself: what’s your purpose for self-hosting? For example, my primary use case is to host two models for coding - one for agentic work (Qwen3-Coder-Next-80B) and one for inline suggestions (Qwen3-Coder-30B-A3B). I also have gpt-oss-20b on standby, because it suits my wife’s workflows very well, but she only uses it sporadically (weirdly, she finds it to be better than ChatGPT).
It’s no different to seeking advice from people, really; you wouldn’t ask a maths professor about Tianenmen Square, but you also wouldn’t ask a historian about the Riemann Hypothesis. For me, choosing the right model for the job at hand is a very important part of the process.
Incidentally, with llama.cpp’s router mode, you can hot-swap models as you need them, and - unlike Ollama - you can set per-model configurations (different context lengths per model etc).
I think I heard somewhere (terrible reference I know so take it for what it’s worth) that US companies are openly lobbying for there to be licensing requirements in order to create an LLM so…
Also, the US government is moving in the direction of banning open source models (from you know where) as a security risk with, I think, fines and jail time attached if you use them.
You could learn how they are trained and decide what training you desire for your output. They are not general conversationists with exceptional ability. It’s predicting the next token. When trained they have many tokens that conclude high probablilty of responding with Nope. This is an option you pick with the trainingset the model used. If you want to have a universal LLM, you’re not in the right space running it at home. If you want freedom to ask anything to any model, the models come with breadcrumbs. It’s like a song in a different language, you can force the singer to speak english, you can only enjoy what they offer.
Are you suggesting pulling the Uno Reverse card on Qwen and pulling the weights out by a sending a massive amount of queries to it and observing the chain of thought and results to feed your own model training? I think that is what US companies are accusing some of the non-US models of doing. You could do that locally. I don’t know what the TOS of these models is, so while I’m not advocating violating these open weight model’s TOS, if it’s allowed, then maybe that’s an option.
More along the lines of: reading the datasets they use that are public, decide if you are okay with that, and then pull a model to use locally. If you want a great coder that can program Tianenmen Square into your code, Qwen might be the wrong choice.
Reversing weights is an option I guess, if you are dead set on that specific base. It’s easier to build a 1B model yourself than worry about all that enough to filter it out.
we’ve got gemma 4 now; more open source than the chinese models while outperforming them on significantly less memory.
it also doesn’t have orwellian censorship build into it which is a plus.
You are just bad at prompting. No offense. Any LLM can serve as a functional search filter engine. Create instructions in a project, call it web search, and win. If you think one LLM is poorly aligned setup projects in them all and test the same prompt with the same instructions.
After testing my use cases…. yes…also again, as others have said, uncensored versions are readily available.