A 1.29 MB control vector gives Bonsai 2 a new refusal-steering option
Secondchancee bundles Prism ML's existing PQ2_0 model with a runtime control vector, claiming fewer refusals without rewriting its ternary weights.
A small sidecar for an existing local model
Secondchancee has published a refusal-steering bundle for Ternary Bonsai 2 27B, pairing the existing model with a roughly 1.29 MB control-vector file. The author promotes it as an uncensored edition that changes behavior during inference rather than permanently editing the ternary weights.
The repository was created on October 17 at 00:11 UTC, and its model artifact was uploaded at 00:13 UTC. Its later activity includes gateway-script updates. Those timestamps establish a new community package, not a new launch of Prism ML’s underlying Bonsai model.
The inspected file listing contains a 7,206,168,928-byte PQ2_0 model and a separate 1,293,184-byte refusal_vector_27b.gguf. The card says the vector subtracts a refusal direction across the decoder layers at runtime. For local-model experimenters, the hook is a configurable behavior adjustment that leaves the stored low-bit model representation in place.
The author claims this reduces refusal preambles while preserving reasoning quality. The card does not supply a comparative evaluation establishing that outcome. Its claims of no quality degradation and 128K context below 10 GB of VRAM remain author claims, not AITrending measurements. Model-file size alone does not establish total memory requirements.
The runtime matters
The community card describes using llama.cpp control-vector support, but the upstream documentation adds a critical compatibility detail: these PQ2_0 files require Prism ML’s llama.cpp fork. Stock llama.cpp does not support the required ternary format and matching activation transform. Users should follow the upstream runtime guide rather than assume that any existing llama.cpp binary can load the package.
The supplied server script also binds to all network interfaces and does not pass an API-key option. It should not be mistaken for a deployment with authentication already configured. This is a community inference-time steering release; it is neither proof of universal refusal removal nor an independently validated improvement to the base model.
