The quantized model Ternary-Bonsai-2-27B-Abliterated has been released

Developers introduced Ternary-Bonsai-2-27B-Abliterated — a specialized version of the model with 27 billion parameters, optimized for operation with a minimal level of failures. The release is targeted at users who require high accuracy when executing queries without excessive censorship.

The key feature of the build is an extremely low failure rate of just 1.0%, with a complete absence of "over-refusals". During the preparation of the model, a procedure called "abliteration" was carried out, aimed at removing undesirable behavioral restrictions. The cost of changing weights during optimization was 0.071%, which allows preserving the original characteristics of the neural network during significant compression.

The model is presented in the GGUF format using the PTQ1_0 quantization method. To launch this version, a specialized fork of llama.cpp from PrismML is required. The technical implementation provides high performance on consumer equipment, achieving an estimate of 13 points in internal tests in 2 hours of data processing.

This release expands the capabilities of local launch of large language models, offering an alternative to standard weights with strict security filters. The use of the GGUF format simplifies integration into existing local execution environments, allowing for efficient distribution of the load on video memory and system RAM when working with a 27-billion architecture.