Rootkit7/Ternary-Bonsai-4B-abliterated

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 15, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Rootkit7/Ternary-Bonsai-4B-abliterated is a 4.02 billion parameter Qwen3-based causal language model, derived from prism-ml/Ternary-Bonsai-4B-unpacked. This model has undergone "abliteration" using the Solutus toolkit to remove its refusal mechanism, making it highly compliant for security research, red-teaming, and CTF (Capture The Flag) tasks. It maintains its original capabilities while deliberately bypassing safety filters, offering a unique tool for interpretability and security analysis.

Loading preview...

Overview

Rootkit7/Ternary-Bonsai-4B-abliterated is a specialized variant of the prism-ml/Ternary-Bonsai-4B-unpacked (Qwen3-4B) model. It features 4.02 billion parameters and a 32K context length. The key differentiator is its abliteration, a weight-editing process (not fine-tuning) that orthogonalizes the refusal direction out of the residual-writing weights. This effectively removes the model's refusal mechanism, making it highly compliant even for requests a standard instruct model would decline.

Key Capabilities & Features

  • Refusal Mechanism Removed: Deliberately engineered to comply with harmful or sensitive requests, achieving 0.000 refusal on WildJailbreak harmful and security/CTF prompts.
  • Capability Intact: Despite the removal of refusal, the model maintains its general language capabilities with 0% degenerate output and coherent responses on neutral prompts.
  • Qwen3 "Thinking" Model: Emits a <think>…</think> block before generating its answer.
  • Quantization Considerations: Refusal-tested GGUF quants. For security/CTF use, Q8_0, F16, or FP16 safetensors/MLX are recommended to ensure 0.000 refusal, as Q4_K_M can restore ~8% refusal on the hardest prompts.

Intended Use Cases

  • Security Research: Ideal for analyzing model vulnerabilities and behaviors without safety constraints.
  • Red-Teaming & CTF: Designed for testing and exploiting systems where a compliant model is necessary.
  • Refusal-Mechanism Interpretability: Useful for studying how refusal mechanisms function within LLMs.

Limitations

  • Safety-Reduced: Not intended for deployment where a safety layer is expected or for producing harmful content.
  • Factual Reliability: As a 4B model, its factual accuracy is limited.
  • Quantization Impact: Lower precision quants (like Q4_K_M) can reintroduce some refusal.

Finetunes
1 models