A quantized version of the MiniMax-M3 model has been released

Developers presented MiniMax-M3-uncensored-GGUF — an uncensored version of a major language model optimized to run through the llama.cpp library. The original MiniMax-M3 architecture has a total parameter volume of 427B, which requires significant computational resources for local deployment. To ensure operation on consumer and server equipment, the model was converted to the GGUF format using various quantization levels. Users have compression options from IQ2_S to Q8_0. The most compact option in the line is the IQ4_XS version, which is 189.13 GiB in size. The quantization process allowed adapting the model weights for efficient use of RAM while preserving the original logic. Using the GGUF format ensures compatibility with current versions of llama.cpp, allowing the load to be distributed between CPU and GPU depending on available hardware capabilities. This release is aimed at researchers and enthusiasts who need access to unfiltered model responses with 427B parameters in a local environment. The presence of a wide range of quantized versions allows finding the optimal balance between generation accuracy and the amount of occupied video memory or RAM.