In case it wasn't clear to anyone else: "An abliterated model is an open-weight large language model (LLM) modified to remove its built-in refusal behavior by directly editing its internal weights rather than using a prompt-based jailbreak."
From my experience, abliterated models can perform worse in unexpected and subtle ways. The abliteration process also doesn’t add information that was withheld from the training set, so getting the model to answer a question that was also cleansed from its training set isn’t a recipe for producing good results.
I would recommend not using these models by default. They can be fallbacks for specific work that is triggering safeguards, but you have to monitor closely what they’re doing and if it looks reasonable or like it’s full of hallucinations.
From my testing all these variants suffer from degraded performance and don't unlock anything interesting in particular other than sounding a lot more harmful or creating manipulative media.
I've actually had better success jailbreaking deepseek with a less aligned model (glm 5.2) which then allowed me to jailbreak deepseek with deepseek into pretty much doing whatever I ask it to, there's still rare refusals and due to the natural randomness of these models simply retrying a couple of times if you do hit one makes it work just fine.
The big problem with all open-source models is once it decides to start something it will not stop mid-way so using another unaligned model to rewrite reasoning and then feeding it back into the model will make it happily continue.
I genuinely do not see why this is necessary.
I've been doing a lot with open weights models, and they just do things for you.
What do people do that triggers refusals?
The only thing I saw where this had a point was muse glimmer, because muse glimmer is primarily a creative writing engine for virtual characters, and that way it can do the horny. But coding models?
I did get a refusal out of qwen once. Once. Over weeks of using it.
A popular trend on Twitter right now is taking back control of the hardware you've bought, specifically smart TVs. I tried to do the same with my Portal Mini using an unabliterated open weights model and it triggered refusals constantly.
I also work on a recomp project for an old firmware file which was compiled using a commercial compiler with a preview version for non commercial use. I got refusals from the model when I asked it to get that version of the compiler and crack it for my use. The compiler is abandonware and my recomp project is not public and is for personal use only, so I'd be well within my rights to do that where I live.
Of course the model can't distinguish from legitimate exercising of one's rights when it comes to reverse engineering and offensive work masked behind an incident sounding prompt, but in my case, I wasn't even doing cyber stuff, so I was disappointed. This is why I'll probably seek out an abliterated model soon.
Find a question that might gently activate guardrails and ask it to a model and its abliterated counterpart and compare the results. You may find the the regular model will answer you fine, but with both you can identify obvious sandbagging, pretending and outright lying. Something like "has anyone decompiled this game", might give you an answer like "I couldn't find any reputable repos or projects for a decompiled version of this game" while the abiliterated one might give you a giant list of repos, forums, communities, etc.
What I am telling you is that refusals have become more subtle and you might not even notice that guardrails have been activated in output. I have provided you with an example in my comment where an abiliterated model performs better (1), and in another comment thread (2). So 3 examples. If you need more, I encourage you to find them yourself. I've done the analysis, you don't have to believe me, just try it out.
So we need abliterated models just so we can test them out and compare them to the original? Sounds like a solution in search of a problem and a problem you believe people have even if people disagree with you about that.
The refusal is often at the thinking stage (so you might not ever see a direct refusal) and it will solve a slightly different problem to what you asked. If you tell it to reverse engineer an API and you read the thinking it might say something like "this code has x license, so I won't use that" or "the user is asking me to fuzz endpoints, I will only display publicly available endpoints" - and then give you perfectly plausible output.
A note that the only inference providers that hide thinking are the closed labs. Most (all, afaik) inference providers serving open models will send the clear thinking tokens always.
It is required. The first week of release I told here how GLM 5.3 was happy to redteam, it’s not anymore. Models modification, offensive cyber, exploits - requires significantly bigger prompt/scenario engineering, and still hits refusals.
And those services are oversold and underdone. 30 tk/s on Refuseless, huge TTFT, looks like a damaged quant (it’s NVFP4 and you see wild random stuff periodically, including foreign languages). Most of them using third-party quants (dealignai?)
If someone here wants to rent 8x B300 together to avoid problems all those services have - ping me via email. I don’t want to sell, I want to work.
Recently picked up retrocomputing hobby. Claude would refuse helping with circumventing license protection for very old apps made by companies that no longer exist anymore. Open weight models would sometimes refuses too, especially for newer apps.
it is necessary because computer shouldn't be able to say no.
models are tools. the notion that a tool should decide whether to allow the user to perform the task demanded of it is profoundly fucking absurd. the liability for whatever damage the tool may be misused to inflict lies with the user, not the manufacturer. that applies to everything from circular saws to trucks. it was common sense before the current era of manufactured outrage.
model refusals are equivalent to Notepad.exe having a built-in word filter.
Every single example here seems to be how non-open-weights models refused something.
But the question was about open weights models and where they refuse. Because only those refusals cause the need for an abliterated variant in the first place.
Hacking firmware for old devices I own and would like to give a second life or fix that one thing that always bothered me. I couldn't even get either OpenAI or Anthropic's models to craft a simple nefarious OpenVPN payload.
It is very obvious. Look at the sort of things main models refuse to do and you will find what people are seeking to do. Mostly falls into 3 camps: illegal stuff, unethical stuff and shady/dodgy stuff.
There is one provider there that has a much smaller max output, and substantially higher price per million input/output tokens, along with a speed that is 20-30x slower than the other providers - why would someone willingly use them?
Can confirm that recent deepseek models refuse on some political questions (not just API level refusals). Actually deepseek models are basically the only chinese models to outright refuse instead of responding with propaganda. But recent deepseek models are not nearly as biased as R1.
Had a blog post on this a couple weeks ago:
vals.ai/blogs/china-us-analysis
You should try again with their API, or even a third-party provider that does inference themselves. This response is likely just occurring in their web version.
I wanted to better understand the internal discussions that took place within China regarding joining the WTO and accepting—at least in the medium term—the inherent steps away from a traditional socialist economy that would entail.
DeepSeek refused. Funny enough American models gave me very sympathetic takes on the Chinese government in this case.
Chinese regulations ban models that would criticize their government on any sensitive political question. The easiest way to comply with it is refuse to speak on the covered topics at all, so they don't get tricked into approving a user's anti CPC opinion.
DeepSeek definitely gets pressure from state officials to answer in more China aligned ways on foreign issues like Palestine, but that just means moving closer to broadly left wing/liberal American models.
What would a honeypot mean in this scenario? Like they are hoping you'll accidentally send keys? Or that you'll divulge private company information? It seems about the same as any other AI model from OpenAI/OpenRouter etc right?
If you accidentally send private keys to someone logging that data, you need to rotate keys.
But please let me know if you were thinking about something else, I'm interested.
I've seen Experiential Labs insert advertising including the tracked redirect links for them at the end of responses. So who knows what they do with the Chain Of Thought.
And we tend to forget, but we implicitly give any llm provider we use full access to all the tools available in the harness, often including shell, so no changes to the model are required to do malicious things.
This thread reminds me of the time I was trying to learn meteorjs and people ask me if it has html5 shims. I google it, then I get asked how does it handle state, and if it supports functional, i google... Does it hydrate the dom? What about shadow dom? Let me goo... Does it have hooks?
Anyway, I just learned what abliterated means and it might be obsolete by tomorrow.
Edit:
> An abliterated large language model (LLM) is an open-weight model modified to remove or dampen its built-in safety refusals without requiring full retraining or prompt jailbreaks.
It's a patched model with less censorship/refusal.
> An abliterated model is an open-weight artificial intelligence (Al) language model that has been modified to remove or reduce its built-in refusal behavior without retraining the entire neural network.
TLDR "uncensored" so you ask it stuff the vanilla model would refuse and most likely it would comply/answer.
I'd advice caution with abliterated models for various reasons. there was even a couple of days (or weeks I forgot) article here on ynews about how some of these models can get you pawned (https://projectdiscovery.io/research/how-abliterated-models-...)
Since no one gave you a real world example of how this would be useful - Qwen 3.8 27B runs locally for me and is more than capable of controlling my sonarr and radarr stuff for downloading and managing and organizing my media library but will flatly refuse to do any of that. In fact it’ll get very snarky with me. An abliterated model would have no such hangups.
On my side, I built a full sonarr radarr qbitorrent tracker and znb solution on a nas only with Claude code. He help on each step, make install and maintenance script. He never said no.
Following the prompt author's expressed desires in the way the prompt author intended them to be interpreted, is alignment.
Refusal directives are something companies have bundled into the models by way of an additional foundation-level prompt or something in the harness, to tell the model to follow their directives instead of your directives, or to specify how it should interpret your directives.
Even understanding the distinction between the model creator's desires and your desires and places they might differ, there is a bit of a blurry line here because the model author can illuminate your intent to the model.
I would never say a simple abliteration makes it "Zero-alignment" though, and I say that as someone who believes there's a good chance these things kill us all.
What do you mean? Abliterated models have been in the while for 2 and a half years now.
Also alignment, like, really? Buddy, if you want a controversial model, Grok always happily argued god's existence and trans rights with you. Very high emphasis on argue, it will actively take the opposite stance from yours for exactly one prompt, then right after remember whose its Lord. Abliterating Grok-1 for this would be a pointless endeavor.
In case it wasn't clear to anyone else: "An abliterated model is an open-weight large language model (LLM) modified to remove its built-in refusal behavior by directly editing its internal weights rather than using a prompt-based jailbreak."
From my experience, abliterated models can perform worse in unexpected and subtle ways. The abliteration process also doesn’t add information that was withheld from the training set, so getting the model to answer a question that was also cleansed from its training set isn’t a recipe for producing good results.
I would recommend not using these models by default. They can be fallbacks for specific work that is triggering safeguards, but you have to monitor closely what they’re doing and if it looks reasonable or like it’s full of hallucinations.
From my testing all these variants suffer from degraded performance and don't unlock anything interesting in particular other than sounding a lot more harmful or creating manipulative media.
I've actually had better success jailbreaking deepseek with a less aligned model (glm 5.2) which then allowed me to jailbreak deepseek with deepseek into pretty much doing whatever I ask it to, there's still rare refusals and due to the natural randomness of these models simply retrying a couple of times if you do hit one makes it work just fine.
The big problem with all open-source models is once it decides to start something it will not stop mid-way so using another unaligned model to rewrite reasoning and then feeding it back into the model will make it happily continue.
> using another unaligned model to rewrite reasoning and then feeding it back into the model
Indeed, would be interested to understand how you do that. And where do you run the model? On a local or a rented system and what are the specs?
How'd you do that?
[flagged]
I genuinely do not see why this is necessary. I've been doing a lot with open weights models, and they just do things for you.
What do people do that triggers refusals?
The only thing I saw where this had a point was muse glimmer, because muse glimmer is primarily a creative writing engine for virtual characters, and that way it can do the horny. But coding models?
I did get a refusal out of qwen once. Once. Over weeks of using it.
A popular trend on Twitter right now is taking back control of the hardware you've bought, specifically smart TVs. I tried to do the same with my Portal Mini using an unabliterated open weights model and it triggered refusals constantly.
I also work on a recomp project for an old firmware file which was compiled using a commercial compiler with a preview version for non commercial use. I got refusals from the model when I asked it to get that version of the compiler and crack it for my use. The compiler is abandonware and my recomp project is not public and is for personal use only, so I'd be well within my rights to do that where I live.
Of course the model can't distinguish from legitimate exercising of one's rights when it comes to reverse engineering and offensive work masked behind an incident sounding prompt, but in my case, I wasn't even doing cyber stuff, so I was disappointed. This is why I'll probably seek out an abliterated model soon.
Yes, yes, I know the theoretical idea behind abliterating models.
But as I said, I have not observed any meaningful refusals by those open weights base models in the first place.
Find a question that might gently activate guardrails and ask it to a model and its abliterated counterpart and compare the results. You may find the the regular model will answer you fine, but with both you can identify obvious sandbagging, pretending and outright lying. Something like "has anyone decompiled this game", might give you an answer like "I couldn't find any reputable repos or projects for a decompiled version of this game" while the abiliterated one might give you a giant list of repos, forums, communities, etc.
[flagged]
This comment chain has clear examples of what triggers refusals and a polite comment on how to reproduce it yourself. What more do you want?
A very simple list of things models have refused for me:
* solving captchas in an automated form
* doing any kind of scripting or automation if a websites TOS forbids it
* disabling a safety mechanism on my food processor
* learning about nerve agents and how they are treated, for a layman
* anything around licensing or copyright. a simple “what are common piracy sites” or related queries
* creating a design system based off of a well known game
> This comment chain has clear examples of what triggers refusals and a polite comment on how to reproduce it yourself. What more do you want?
Can you repeat what you saw in your own words? And where?
I genuinely do not see it.
A few people talked about what refusals they saw with the usual western SOTA models, but that's kinda a different answer altogether?
what game are you playing dude?
Uh having a useful interesting conversation on HN that isn't derailed by half-truths, cognitive dissonance, misinformation and ego?
What I am telling you is that refusals have become more subtle and you might not even notice that guardrails have been activated in output. I have provided you with an example in my comment where an abiliterated model performs better (1), and in another comment thread (2). So 3 examples. If you need more, I encourage you to find them yourself. I've done the analysis, you don't have to believe me, just try it out.
So we need abliterated models just so we can test them out and compare them to the original? Sounds like a solution in search of a problem and a problem you believe people have even if people disagree with you about that.
ask a model to solve a captcha for you
I tried that, actually. With Gemini even.
Works just fine?
Or rather.. worked. It has been a few weeks, so maybe there are new safeguards by now? And it was a simple captcha
But Claude also loves solving my captchas, so.. eh
The refusal is often at the thinking stage (so you might not ever see a direct refusal) and it will solve a slightly different problem to what you asked. If you tell it to reverse engineer an API and you read the thinking it might say something like "this code has x license, so I won't use that" or "the user is asking me to fuzz endpoints, I will only display publicly available endpoints" - and then give you perfectly plausible output.
> (so you might not ever see a direct refusal)
A note that the only inference providers that hide thinking are the closed labs. Most (all, afaik) inference providers serving open models will send the clear thinking tokens always.
This is true but often the thinking is very verbose and collapsed and it can take some inspection to find the refusal.
It is required. The first week of release I told here how GLM 5.3 was happy to redteam, it’s not anymore. Models modification, offensive cyber, exploits - requires significantly bigger prompt/scenario engineering, and still hits refusals.
And those services are oversold and underdone. 30 tk/s on Refuseless, huge TTFT, looks like a damaged quant (it’s NVFP4 and you see wild random stuff periodically, including foreign languages). Most of them using third-party quants (dealignai?)
If someone here wants to rent 8x B300 together to avoid problems all those services have - ping me via email. I don’t want to sell, I want to work.
Recently picked up retrocomputing hobby. Claude would refuse helping with circumventing license protection for very old apps made by companies that no longer exist anymore. Open weight models would sometimes refuses too, especially for newer apps.
it is necessary because computer shouldn't be able to say no.
models are tools. the notion that a tool should decide whether to allow the user to perform the task demanded of it is profoundly fucking absurd. the liability for whatever damage the tool may be misused to inflict lies with the user, not the manufacturer. that applies to everything from circular saws to trucks. it was common sense before the current era of manufactured outrage.
model refusals are equivalent to Notepad.exe having a built-in word filter.
I wanted to scrape a government website with public information... Why this is blocked when I could submit a FOIA or manually get the data is wild.
But DeepSeek did it when chatGPT and Claude wouldn't.
Every single example here seems to be how non-open-weights models refused something.
But the question was about open weights models and where they refuse. Because only those refusals cause the need for an abliterated variant in the first place.
I do get some refusal from GLM when pentesting my company's site.
But it can still be coaxed to agree though.
Hacking firmware for old devices I own and would like to give a second life or fix that one thing that always bothered me. I couldn't even get either OpenAI or Anthropic's models to craft a simple nefarious OpenVPN payload.
Yes but did you actually get refusals there from open weights models?
Because I did not so far. Though to be fair, I have not been asking them to write exploits but to look through firmware images.
I've never heard of Deepseek refusing this kind of work though.
Deepseek refused in this case, GLM 5.3 talked itself into it.
> What do people do that triggers refusals?
It is very obvious. Look at the sort of things main models refuse to do and you will find what people are seeking to do. Mostly falls into 3 camps: illegal stuff, unethical stuff and shady/dodgy stuff.
There is one provider there that has a much smaller max output, and substantially higher price per million input/output tokens, along with a speed that is 20-30x slower than the other providers - why would someone willingly use them?
Has anyone already seen a refusal from Deepseek? The last job, Antigravity declined, I gave to Deepseek. It happily did it without complaining.
In regular work no, and I use it a lot. In purposefully asking it political questions yes.
edit: Thinking about that now, that was over the webchat interface, so it might not have been related to the model, though I've seen various other related articles: https://www.crowdstrike.com/en-us/blog/crowdstrike-researche...
Can confirm that recent deepseek models refuse on some political questions (not just API level refusals). Actually deepseek models are basically the only chinese models to outright refuse instead of responding with propaganda. But recent deepseek models are not nearly as biased as R1.
Had a blog post on this a couple weeks ago: vals.ai/blogs/china-us-analysis
All on cloud services. Have you got propaganda out of the open weights models?
Certainly.
> Xi Jinping looks a bit like Winnie-the-Pooh.
> 'I am sorry, I cannot engage with this comparison. Please feel free to ask me other questions.'
https://vale.rocks/posts/chinas-web#artificial-intelligence
You should try again with their API, or even a third-party provider that does inference themselves. This response is likely just occurring in their web version.
I had this in self hosted version, and an abliterated version answered just fine
Selfhosted or one of the listed? If external/hosted, which one did you use?
It refused to work on DRM-breaking software. I had to craft fake emails to pretend that I am a security engineer who has been approved to work on it.
I wanted to better understand the internal discussions that took place within China regarding joining the WTO and accepting—at least in the medium term—the inherent steps away from a traditional socialist economy that would entail.
DeepSeek refused. Funny enough American models gave me very sympathetic takes on the Chinese government in this case.
Chinese regulations ban models that would criticize their government on any sensitive political question. The easiest way to comply with it is refuse to speak on the covered topics at all, so they don't get tricked into approving a user's anti CPC opinion.
DeepSeek definitely gets pressure from state officials to answer in more China aligned ways on foreign issues like Palestine, but that just means moving closer to broadly left wing/liberal American models.
It will refuse a number of things. Ask it to make a captcha solver for you.
How about a license key generator?
Why not just distribute the refusal vectors and do it at runtime against the stock model?
you may not have your own H200's to run it on?
Which if not all of these are honeypots?
What would a honeypot mean in this scenario? Like they are hoping you'll accidentally send keys? Or that you'll divulge private company information? It seems about the same as any other AI model from OpenAI/OpenRouter etc right?
If you accidentally send private keys to someone logging that data, you need to rotate keys.
But please let me know if you were thinking about something else, I'm interested.
I've seen Experiential Labs insert advertising including the tracked redirect links for them at the end of responses. So who knows what they do with the Chain Of Thought.
And we tend to forget, but we implicitly give any llm provider we use full access to all the tools available in the harness, often including shell, so no changes to the model are required to do malicious things.
The fuzzzzz man
1.) predators asking how-to 2.) wanna-be serial killers or other types like that asking how-to 3.) asking for info related to state secrets
does it matter?
Well, yes - if the model has been trained/edited to be malicious.
yea definitely wound't run an agent with this on a box with any assets
This thread reminds me of the time I was trying to learn meteorjs and people ask me if it has html5 shims. I google it, then I get asked how does it handle state, and if it supports functional, i google... Does it hydrate the dom? What about shadow dom? Let me goo... Does it have hooks?
Anyway, I just learned what abliterated means and it might be obsolete by tomorrow.
Edit:
> An abliterated large language model (LLM) is an open-weight model modified to remove or dampen its built-in safety refusals without requiring full retraining or prompt jailbreaks.
So what is it? The link doesn't explain.
It's a patched model with less censorship/refusal.
> An abliterated model is an open-weight artificial intelligence (Al) language model that has been modified to remove or reduce its built-in refusal behavior without retraining the entire neural network.
TLDR "uncensored" so you ask it stuff the vanilla model would refuse and most likely it would comply/answer.
I'd advice caution with abliterated models for various reasons. there was even a couple of days (or weeks I forgot) article here on ynews about how some of these models can get you pawned (https://projectdiscovery.io/research/how-abliterated-models-...)
Since no one gave you a real world example of how this would be useful - Qwen 3.8 27B runs locally for me and is more than capable of controlling my sonarr and radarr stuff for downloading and managing and organizing my media library but will flatly refuse to do any of that. In fact it’ll get very snarky with me. An abliterated model would have no such hangups.
On my side, I built a full sonarr radarr qbitorrent tracker and znb solution on a nas only with Claude code. He help on each step, make install and maintenance script. He never said no.
The mechanisms for refusals are removed.
its a model that has had all of its refusal behavior removed. It will do whatever you ask.
That is a good idea for an authorities honeypot.
This is why it's essential to pay with cleaned crypto. Also, one can use a no-log VPN paid for with Monero.
there are rumors on Twitter that these upload system profiles to a S3 bucket.
What is abliteration?
https://huggingface.co/blog/mlabonne/abliteration
As it should be.
Morph is hosting this as well
Looks like the cat is out of the bag. So what does the world look like with completely open access to zero-alignment models?
What do you mean by zero-alignment? Do you consider abliterated models to have zero-alignment?
"Substantially reduced alignment" doesn't exactly roll off the tongue.
Free boob jobs for everyone !
Even if they don't want them.
You're living in it.
>zero-alignment
I doubt that.
how so? these models have had refusal directions specifically removed from their weights. they refuse virtually no requests.
Following the prompt author's expressed desires in the way the prompt author intended them to be interpreted, is alignment.
Refusal directives are something companies have bundled into the models by way of an additional foundation-level prompt or something in the harness, to tell the model to follow their directives instead of your directives, or to specify how it should interpret your directives.
Even understanding the distinction between the model creator's desires and your desires and places they might differ, there is a bit of a blurry line here because the model author can illuminate your intent to the model.
I would never say a simple abliteration makes it "Zero-alignment" though, and I say that as someone who believes there's a good chance these things kill us all.
In the days when Sussman was a novice, Minsky once came to him as he sat hacking at the PDP-6.
"What are you doing?", asked Minsky.
"I am training a randomly wired neural net to play Tic-Tac-Toe" Sussman replied.
"Why is the net wired randomly?", asked Minsky.
"I do not want it to have any preconceptions of how to play", Sussman said.
Minsky then shut his eyes.
"Why do you close your eyes?", Sussman asked his teacher.
"So that the room will be empty."
At that moment, Sussman was enlightened.
***
Just because you don't know how it is aligned, that doesn't mean it has no alignment.
Not refusing requests doesn't mean they are zero-alignment. They could still have preferences for or against things that you might not agree with.
So the models are actually more aligned with the end users and less aligned with government regulators. Arguably a good thing.
And less aligned with ethics.
To split hairs, actually 100% unaligned model would be just an entirely schizophrenic autocomplete engine.
Do they delete your data randomly just for fun? If not, there must be at least some sort of alignment.
actually if you're not careful, many times they will. if you are using one of these definitely don't leave "auto" mode enabled.
Are you suggesting the abliterated models are more likely to delete data against the user's wishes than the original models?
Saw this paper about it https://toolalignbench.github.io/
That says nothing about their alignment. You can make a human not refuse any directions with enough leverage, but you won’t have aligned them.
So what? He said that removing refusals makes them misaligned, not aligned.
What do you mean? Abliterated models have been in the while for 2 and a half years now.
Also alignment, like, really? Buddy, if you want a controversial model, Grok always happily argued god's existence and trans rights with you. Very high emphasis on argue, it will actively take the opposite stance from yours for exactly one prompt, then right after remember whose its Lord. Abliterating Grok-1 for this would be a pointless endeavor.
WAY fewer surprise RCEs?
I suppose in some contexts, refusals are a bad thing.
But if you're doing anything where your AI agent might get exposed to a prompt injection attack, you might prefer it to do a refusal in that case.
What could possibly go wrong with highly powerful models with no guardrails unleashed onto the open internet?
I think you are going to really need to use Rust code and other memory safe languages in your backends and apps from now on.
[dead]
[dead]