Ternary Bonsai 2 27B Steered

Secondchancee’s October 17 community package distributes an existing Bonsai 2 PQ2_0 model with a separate refusal-steering vector. The author describes changing model behavior during inference without rewriting the ternary weights. Claimed refusal reduction, quality preservation and memory usage have not been independently reproduced by AITrending.

The underlying PQ2_0 format requires Prism ML’s llama.cpp fork, according to its upstream model guide. The supplied community server script binds beyond loopback without passing an API-key flag; authentication and network access need separate attention before remote serving.