0xzknw/LFM2.5-1.2B-Thinking-Heretic

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.2BQuant:BF16Context Size:32kPublished:Aug 24, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

0xzknw/LFM2.5-1.2B-Thinking-Heretic is a 1.2 billion parameter experimental derivative of LiquidAI's LFM2.5-1.2B-Thinking model, developed by 0xzknw. This model has been 'abliterated' to attenuate internal refusal directions while largely preserving benign prompt behavior. It achieves a refusal marker score of 3/100 and a KL divergence of 0.0003 on benign prompts, making it suitable for use cases where reduced refusal behavior is desired.

Loading preview...

LFM2.5-1.2B-Thinking-Heretic Overview

This model is an experimental 1.2 billion parameter derivative of LiquidAI's LFM2.5-1.2B-Thinking, developed by 0xzknw. It has been modified using an 'abliteration' method to significantly reduce its tendency to refuse prompts, while aiming to maintain its original behavior on benign inputs. The modification process involved estimating per-layer refusal directions and projecting output weights away from these directions.

Key Characteristics & Performance

  • Refusal Attenuation: Achieves a refusal marker score of 3/100 (compared to 98/100 for the original checkpoint) on a set of 100 refusal-oriented prompts.
  • Benign Behavior Preservation: Maintains a low KL divergence of 0.0003 on 100 benign prompts, indicating minimal deviation from the base model's responses in non-refusal scenarios.
  • Architecture: Adapted for LFM2.5's hybrid architecture, which includes 10 convolutional LIV blocks and 6 GQA blocks.
  • Availability: Provided in native Transformers safetensors (BF16) and GGUF (BF16) formats for compatibility with llama.cpp and LM Studio.

Important Considerations

  • Safety: Abliteration deliberately weakens refusal behavior, which may remove useful safeguards and increase harmful compliance. This checkpoint is not safety-aligned and requires thorough evaluation for specific use cases.
  • Base Model Behavior: Inherits the base model's tendency to spend more than 512 tokens in <think> on simple instructions.
  • License: Distributed under the LFM Open License v1.0, which includes a commercial-use revenue threshold.

Finetunes
2 models