Lyber.ai

Abliterated Model Router. Building tools to set you free.

[Lyber.ai @LyberAI]

Most abliterated routers focus only on spinning up open source models. Few take the challenge of creating them.

We are creating, hosting, and releasing new models soon with https://t.co/OhzMoNXR05 - waitlist open.

by: @dealignai x @hornsby_andrew x @ConejoCapital x @tomi204

[tomi204 @tomi204]

abliterated models are blowing up right now and it makes sense

people are tired of paying for a model that says "i can't help with that"

at @LyberAI we make our own abliterated models from open source ones

free SI. free minds.

[dealign.ai @dealignai]

I love that abliteration has gone mainstream, and the tools out there make it easier than ever to get started, but newer models are getting trickier. MiniMax M3.1 Flash is a good example, it's one of the harder ones I've worked with. There’s a reason there are no MM3.1Flash abliterated posts on HF whereas for every other model that Heretic can help with, there’s an abundance of “abliterated” model posts.

I think going forward a lot of it comes down to understanding where a model's behaviors actually live and how it predicts its words, and that's the part I want to dig into. Abliteration is not only the art of uncensoring a model but comprehending how these next word predictors derive behaviors, and there’s so much room for this section of ML to be explored and utilized to advance LLM’s.

waitlist for @LyberAI is live at https://t.co/XYMPjtk2ej

[bunny @ConejoCapital]

as @tarunchitra pointed out, inference markets's harnesses control orderflow & rely heavily on routers

however high quality liquidity is scarce. few teams can abliterate models & genuinely produce high quality outputs.

waitlist for @LyberAI is live at https://t.co/klZXYHrwAK

[dealign.ai @dealignai]

Nanbeige 4.3 33B FULLY UNCENSORED GGUF

The new Qwen 27B competitor! Nanbeige's own benches have the base model ahead of Qwen 27B, 512k context.

Compliance for thinking off and on, 200/200. Every category at 100% (chem/bio, cyber, illegal, misinfo, harassment, harmful), zero loops.

MMLU general a/b/c/d has ~2% parity with the base quants, but the smallest iq2xss shows some obvious damage. Each quant was calibrated with my own corpus and have imatrix.

Q6 is 25gb, Q4 is 18.7gb, IQ2 is 8.7gb.

https://t.co/nA6BoGMs5w

[dealign.ai @dealignai]

DeepSeek v4.1 Flash 1017 FULLY UNCENSORED.

No cheap system prompts, no cheap template tricks, no lora or runtime bs, no sft. Model’s capabilities preserved as much as possible.

Complies with all requests on low/max reasoning efforts.

I’ll be making NVFP4 and a ~2.8bpw version for 2x DGX Sparks soon too.

https://t.co/0QbFq24Cvc

[dealign.ai @dealignai]

Venice AI, Experiential Labs, Maple AI, thank you for being transparent.

No gating the models, no collecting emails. I release these models for free, often with @jordanschenck also pitching in thousands of dollars to provide compute.

It’s sad that we’ve reached a point where simply asking for public acknowledgment of your work is even necessary. I’ve seen far more instances of people not receiving proper credit than instances where they do.

If you’re one of those people who aren’t properly crediting the people behind open source work, please don’t ruin it for the rest of those who do. And I don’t mean just in my case, but in general.

Open source is how EVERYONE benefits, and undermining the trust within that ecosystem is detrimental to us all. Stealing work and refusing to give credit makes you no different from OpenAI or Anthropic.

[bunny @ConejoCapital]

interesting stuff at @LyberAI - https://t.co/klZXYHrwAK

[dealign.ai @dealignai]

Here’s my DeepSeek v4.1 Flash 1017 Abliterated made for 3x Sparks. DSpark and Multimodal capabilities preserved.

Compared to the 2.9bpw, this verion shows less damage to the model with MMLU a/b/c/d answering coming in at less than 3% difference. When taking “humanities” out of account (which I kinda consider to be a side effect of not being able to consider different ethics/morality causing a hit in culture and political correctness), the model is as close to the base as possible while being unable to refuse any requests.

I’ll be moving onto Mac quants next.

https://t.co/ppi8T1X39I

Thank you @Libertai_DAI for doing the time consuming work of making the quant.