Uncensored MiniMax Model Cut from 426B to 112GB Released by OrcaRouter

At a glance

OrcaRouter has released an uncensored version of the MiniMax M3.1 Flash model in GGUF format. It says the original model, which is around 426B in size, was compressed to 112GB so it can run on one computer with 128GB of memory. It says refusals were reduced while most of the model's abilities were kept.

Key points

1. 1

   It is a build in GGUF (a file format read by programs in the llama.cpp family).

2. 2

   Abliteration (a process that removes the direction in which the AI refuses requests) was applied.

3. 3

   When run directly, prompts and answers do not leave your own computer.

4. 4

   The same model can also be used through an API, but that version is the full-precision version without compression.

Why it matters

High-performance AI can now run on personal hardware, and because the model has fewer safety guardrails, it can be useful for research and security testing, which makes it significant.