The quantized model Qwen3.8-27B with MTP optimization has been released

Developers presented Qwen3.8-27B-abliterated in GGUF format, optimized for local execution. The model is based on the AEON-7 architecture and uses the Ridge-inspired mixed quantization method, which allowed the weights to be fitted into 12 GB of video memory.

The technical characteristics of the build include a compression level of 3.69 bpw. When tested on the RTX 5060 Ti video card, the model demonstrates a generation speed of around 37 tokens per second. The toolkit is adapted to work through the llama.cpp engine, which provides compatibility with most consumer GPUs.

In internal tests, the model scored 106 points based on the results of a one-hour run. The use of the MTP structure (Mixed-quantization Tensor Packing) allows minimizing the loss of accuracy while significantly reducing the requirements for the amount of VRAM, while maintaining high inference speed on local machines.

This release is aimed at users who need to run heavy models with 27 billion parameters on mid-range equipment. Optimizing the weights to 12 GB makes the model accessible to systems with limited memory without the need to use cloud resources.