UNCENSORED AI 2026 - RAG capabilities

[LeviAckkerman]

Hey, I am looking for a completely uncensored AI that accepts being trained on files (RAG), as far as I know the terms. It could be offline/online. I need good recommendations. If you know something good, let me know.
I tried those:
1. Stansia AI - yeah, it's uncensored, but sometimes I feel it lacks. (online)
2. Venice AI - I didn't try this one; I know it's online, but maybe you have an opinion here.
3. Dolphin/Qwen3 - offline AI, it does the job, not that much creativity, for offline, I think those are the best (from my research).

Or if you have a method to create your own AI-based, maybe on the currently leaked ClaudeCode, or something else, if you could point me in a direction, that would be great!

Thanks!

 

[LeviAckkerman]

bump

 

[Peb]

Hey, I am looking for a completely uncensored AI that accepts being trained on files (RAG), as far as I know the terms. It could be offline/online. I need good recommendations. If you know something good, let me know.
I tried those:
1. Stansia AI - yeah, it's uncensored, but sometimes I feel it lacks. (online)
2. Venice AI - I didn't try this one; I know it's online, but maybe you have an opinion here.
3. Dolphin/Qwen3 - offline AI, it does the job, not that much creativity, for offline, I think those are the best (from my research).

Or if you have a method to create your own AI-based, maybe on the currently leaked ClaudeCode, or something else, if you could point me in a direction, that would be great!

Thanks!

I’ve played around with a few of these setups, and honestly “completely uncensored” + good performance is still a bit of a trade-off right now.

For offline, you’re already on the right track with Dolphin/Qwen-type models. A lot of people also lean toward things like Mistral-based or Mixtral variants with a local setup (usually via something like Ollama or LM Studio). The creativity can feel a bit limited, but you gain way more control and privacy, especially when you plug in your own RAG pipeline.

For RAG specifically, the model matters less than the setup. If you pair a decent local model with something like LangChain or LlamaIndex and a solid embedding model, you can get really good results from your own data. That’s usually where the “custom AI” feel comes from, not just the base model.

On the hosted side, Venice is decent from what I’ve seen people say, more flexible than most mainstream tools, but still not totally unrestricted. Most online platforms will always have some level of filtering.

If you want full control, the best route is probably:
run a local model → hook it into a RAG pipeline → fine-tune prompts/system behavior over time. That gives you way more flexibility than relying on any single platform.

It’s a bit more setup upfront, but once it’s running, it’s way closer to what you’re describing than any plug-and-play tool right now.

 

[LeviAckkerman]

I’ve played around with a few of these setups, and honestly “completely uncensored” + good performance is still a bit of a trade-off right now.

For offline, you’re already on the right track with Dolphin/Qwen-type models. A lot of people also lean toward things like Mistral-based or Mixtral variants with a local setup (usually via something like Ollama or LM Studio). The creativity can feel a bit limited, but you gain way more control and privacy, especially when you plug in your own RAG pipeline.

For RAG specifically, the model matters less than the setup. If you pair a decent local model with something like LangChain or LlamaIndex and a solid embedding model, you can get really good results from your own data. That’s usually where the “custom AI” feel comes from, not just the base model.

On the hosted side, Venice is decent from what I’ve seen people say, more flexible than most mainstream tools, but still not totally unrestricted. Most online platforms will always have some level of filtering.

If you want full control, the best route is probably:
run a local model → hook it into a RAG pipeline → fine-tune prompts/system behavior over time. That gives you way more flexibility than relying on any single platform.

It’s a bit more setup upfront, but once it’s running, it’s way closer to what you’re describing than any plug-and-play tool right now.

Great answer, thank you!

 

[Gogol]

Hey, I am looking for a completely uncensored AI that accepts being trained on files (RAG), as far as I know the terms. It could be offline/online. I need good recommendations. If you know something good, let me know.
I tried those:
1. Stansia AI - yeah, it's uncensored, but sometimes I feel it lacks. (online)
2. Venice AI - I didn't try this one; I know it's online, but maybe you have an opinion here.
3. Dolphin/Qwen3 - offline AI, it does the job, not that much creativity, for offline, I think those are the best (from my research).

Or if you have a method to create your own AI-based, maybe on the currently leaked ClaudeCode, or something else, if you could point me in a direction, that would be great!

Thanks!

I am personally using the qwen 3 coder abliterated version:

https://ollama.com/huihui_ai/qwen3-coder-abliterated:30b

As for RAG, try using LM studio:

https://lmstudio.ai/

It has RAG feature built in (and ofcourse you can use Qwen 3 directly from LM studio).

Also try using Block Goose editor if you want to use Qwen 3 coder to directly alter the files for you (like google antigravity for example). Here's the project if you are interested:

https://github.com/aaif-goose/goose

The biggest problem with the local AI (atleast with the m1 max mackbook pro I have) is when you need to edit an existing project. The token size in this case becomes really big and the agents start failing. But for a new project from scratch, my setup has worked wonderfully well.

P.S.: I am not affiliated to any of the above mentioned projects.

 

[r2150523]

The biggest problem with the local AI (atleast with the m1 max mackbook pro I have) is when you need to edit an existing project. The token size in this case becomes really big and the agents start failing. But for a new project from scratch, my setup has worked wonderfully well.

It doesn't work on my laptop. It says "8 gb ram required, you have 4 gb".
Lame stuff!

Anyone has seen uncensored model in cloud?

1. Stansia AI - yeah, it's uncensored, but sometimes I feel it lacks. (online)
2. Venice AI - I didn't try this one; I know it's online, but maybe you have an opinion here.

Okay, got it.

 

[LeviAckkerman]

It doesn't work on my laptop. It says "8 gb ram required, you have 4 gb".
Lame stuff!

Anyone has seen uncensored model in cloud?


Okay, got it.

Amazing profile pic btw haha

 

[Gogol]

It doesn't work on my laptop. It says "8 gb ram required, you have 4 gb".
Lame stuff!

You do need a bit of RAM if you want to run local AI. You could also run something like ollama on a VPS, but I have not personally tested it. May be setup a linode server and try installing ollama in it.

 

[r2150523]

You do need a bit of RAM if you want to run local AI. You could also run something like ollama on a VPS, but I have not personally tested it. May be setup a linode server and try installing ollama in it.

It's about VRAM I think which I have 4 GB. I have 16 GB DDR5 RAM.

Typical VPS won't have much VRAM.

 

[Gogol]

It's about VRAM I think which I have 4 GB. I have 16 GB DDR5 RAM.

Typical VPS won't have much VRAM.

Hmm it could be. That's one more reason to invest in apple silicon then.

 

[r2150523]

Hmm it could be. That's one more reason to invest in apple silicon then.

Which Apple device model? And what uncensored model locally it would run?

I know these devices are good, energy consumption is extremely low and you can sit outside your house or office with it.

I don't have it, but I can get it.

 

[Gogol]

Which Apple device model? And what uncensored model locally it would run?

I know these devices are good, energy consumption is extremely low and you can sit outside your house or office with it.

I don't have it, but I can get it.

I have the m1 max macbook pro with 32gb RAM. The newer versions are much, much better if you have the budget. If not, this will be enough.

As for the model, I am running the qwen 3 coder abiliterated version (check my first post for the link).

 

[r2150523]

I have the m1 max macbook pro with 32gb RAM. The newer versions are much, much better if you have the budget. If not, this will be enough.

And what local AI model can you run on it? If any...

 

[Gogol]

And what local AI model can you run on it? If any...

https://ollama.com/huihui_ai/qwen3-coder-abliterated:30b

This runs fine on the m1 max. Here's my settings:


I did try maxing out the token size but it doesn't work properly after around 80K tokens. May be if you buy the 64G RAM version of macbook pro, or the high end mac studio machines, you can get even more.

 

[r2150523]

https://ollama.com/huihui_ai/qwen3-coder-abliterated:30b

This runs fine on the m1 max. Here's my settings:
View attachment 517437

I did try maxing out the token size but it doesn't work properly after around 80K tokens. May be if you buy the 64G RAM version of macbook pro, or the high end mac studio machines, you can get even more.

No way. I will run in cloud because I don't have any income from it.

But I think for video this machine would be good. Not worth for 30B qwen.

However, RTX GPUs are better?? I don't know.

 

[Gogol]

But I think for video this machine would be good.

It is pretty good for video, that's for sure. I have used davinci resolve on this machine, and it ran beautifully (that too on battery power). I have only done 1080p videos and the playback / scrubbing works fine even without proxy clips (but that would depend on the source clips to be honest). I have heard from others that 4k projects run just fine on m1 max mbp.

However, RTX GPUs are better?? I don't know.

I mean, if I could get good hardware without having to sell my kidneys, I would buy it right now. Due to the memory and gpu shortage these days, some mac models are actually cheaper to buy if you are comparing hardware. Once this exorbitant pricing comes down a little, I am sure it will be better to build a custom PC with maxed out hardware.

 

[r2150523]

It is pretty good for video, that's for sure. I have used davinci resolve on this machine, and it ran beautifully (that too on battery power). I have only done 1080p videos and the playback / scrubbing works fine even without proxy clips (but that would depend on the source clips to be honest). I have heard from others that 4k projects run fine on m1 max just fine.

Oh no, I meant AI video generation. Not davinci editing.

I mean, if I could get good hardware without having to sell my kidneys, I would buy it right now. Due to the memory and gpu shortage these days, some mac models are actually cheaper to buy if you are comparing hardware. Once this exorbitant pricing comes down a little, I am sure it is will be better to build a custom PC with maxed out hardware.

You make it back in electricity costs. Paradox.

 

[Gogol]

Oh no, I meant AI video generation. Not davinci editing.

I see, I have not tried it sorry. May be someone else here has?

 

[Temperrrrr]

Hey, I am looking for a completely uncensored AI that accepts being trained on files (RAG), as far as I know the terms. It could be offline/online. I need good recommendations. If you know something good, let me know.
I tried those:
1. Stansia AI - yeah, it's uncensored, but sometimes I feel it lacks. (online)
2. Venice AI - I didn't try this one; I know it's online, but maybe you have an opinion here.
3. Dolphin/Qwen3 - offline AI, it does the job, not that much creativity, for offline, I think those are the best (from my research).

Or if you have a method to create your own AI-based, maybe on the currently leaked ClaudeCode, or something else, if you could point me in a direction, that would be great!

Thanks!

I got few good ones