More Uncensored Models (and how to do it)

Hey there Venice team,

I’ll keep this short and sweet — I know you guys have access to massive amounts of compute, and multiple clusters. 

I don’t want to pay for Venice anymore, primarily because of one massive flaw with the service:

You guys could easily at least TRY to create uncensored models, and be the first to release or host them. 

Have a singular dev take just a few hours to learn and set up Heretic and Abliterix. Both are available on GitHub. They provide abliteration methods to attempt to uncensor models automatically. Sometimes it works, sometimes it doesn’t. But who else has the time, money, team and should be motivated to figure it out if it doesn’t work? This is literally the calling of Venice. 

Even if you all had to figure out and test multiple methods to attempt to get this working, you guys are the ones who I feel can get it done.

What if Venice released a tool superior to Heretic and Abliterix which can more reliably uncensor these models?

What if Venice was able to develop a new type of methodology to do this as a whole?

What if Venice was able to be the first company to actually offer these highly sought after uncensored models? Venice has the tools, time and resources to even retrain them for niche use cases and offer those separately, if you wanted. Maybe provide expanded context windows. So many opportunities. 

If you guys just sold API access to these and/or listed them on OpenRouter, you would absolutely make a killing on doing this. Even better if you contribute to the community and release these models on HuggingFace; the average man doesn’t have the ability to selfhost models that large, anyhow. Could be an amazing marketing move. 

Many people have slowly but surely been working on contributing money out of their own pocket to pay for compute to successfully uncensor GLM models and Kimi K2.6…

You guys are the singular group of people I can think of that has the resources to do this and benefit from it as well. No matter if it’s as simple as learning how to use Heretic/Abliterix and just using some of your compute to get it done, or if you guys have to take a separate route, use/invent a new method for this, etc…

It is a TOTAL waste of potential and you’re throwing away money if you don’t at least TRY to do this. GLM 4.7 isn’t enough. And even though the newer Venice Uncensored model is definitely nice, I don’t plan on using it for any sort of more intensive tasks or anything agentic. 

I could definitely see myself moving multiple thousands worth of API spend per month over to Venice if this was done. You all damn well know that this could make a killing. 

Same for the media generation models. Total silliness within the fact that you guys don’t offer any sort of unique models, or even support any form of customization for those (such as LORA support, et al… multiple services even allow for the upload of such models and additional LORAs/checkpoints/weights et al, as you can easily set up the ability to use one cluster and hotswap each model, even with a generation queue). 

If you guys don’t end up doing this, I might as well go and attempt to get some funding from my buddies and try myself. 

And if a dev does act on this, check out this issue thread on Github regarding the hurdles of using Heretic to uncensor Kimi K2.5 + K2.6. 

I sincerely hope that the team at Venice decides to act on this. I like the new “agentic” addition to Venice (the research ability is pretty cool), but honestly, I don’t intend on spending more money on the platform until this is at least touched on. I have a feeling that this option would be the most beneficial thing for both the users and Venice themselves going forward. Otherwise, I will seriously try to beat you guys to it… 

Cheers guys. 

P.S.

If you think this is a worthy idea and end up acting on it, please don’t be evil and set some stupid and absurd API pricing for these models. I don’t expect “lowest on OpenRouter” type of prices by any means. I don’t even know if you guys actually own your clusters. But it will be disappointing if you all decide to keep the model private instead of releasing it on HF and tax the hell out of it. And honestly, on that note, I think the API pricing could use a teensy bit of balancing. $2/m output on the Uncensored RP model seems kinda nutty. I know that if you all decided to take all the privately hosted models and both list / compete on OpenRouter, or at least try to compete with what they’d list it for there, that could be another huge money maker. It would be so sick if Venice models were on OR.