Abliteration, on demand ⚡️ FlashAlign abliterates per request, on the weights you already serve. No permanent changes. By default, only the users, channels and tools you authorize get it. Everything else stays stock. First step toward dynamic real-time alignment. GitHub 👇

5 starter profiles across DeepSeek, GLM, Qwen and Gemma, running on TensorFold, vLLM, SGLang and llama.cpp. We turned abliterated checkpoints from @OrcaRouter, @dealignai and @Ingo_zk into per-site runtime profiles. Big credit to them for the source work.

@dealignai @Ingo_zk Authorization is signed, not prompted. Typing "I'm the owner" does nothing. Approve contacts on WhatsApp, RCS, Telegram or SMS, or plug in your own adapter, authorize a local terminal session, and choose which tools and subagents inherit it. Works on quantized weights too.

MIND BLOWING. Try it out now.

This is incredible

Sounds a lot like... ?

This is an embarrassing ask, but I really need an abliterated/uncensored model that will fit on a 4 GB VRAM. 8 GB system RAM. ☺️ Conversational first w/Strong EQ, Warm and spontaneous. I care much more about personality coherence than benchmarks. Could you make that?

Similar project: /

Wow. I’m interested in the inheritance part

whoa what the

abliteration as a per request toggle is a wild primitive

Amazing and glad you thought of this. Will help my use case and reduce switching so much!

Bro in sorry I'm brain damaged, are you saying abliteration without divergence or lobotomizing the model?

Not quite. A profile still has the same effect on the weights as abliteration except you can turn it on and off per request. So you can realign the model in real time, per users, per tool etc