saidutta69/LFM2.5-1.2B-Thinking-heretic
The saidutta69/LFM2.5-1.2B-Thinking-heretic is a 1.17 billion parameter, decensored variant of LiquidAI's LFM2.5-1.2B-Thinking model, created using Heretic v1.4.0 for directional ablation. This model retains the base's hybrid conv+attention architecture, 32,768 token context length, Pythonic tool-call format, and multilingual coverage, while significantly reducing refusal behavior from 98/100 to 10/100 via targeted weight edits. It excels in reasoning tasks, emitting an explicit trace before its answer, making it suitable for on-device local agents, math, multi-step reasoning, RAG, and structured extraction.
Loading preview...
LFM2.5-1.2B-Thinking-heretic: Decensored Reasoning Model
This model is a 1.17 billion parameter, decensored variant of the LiquidAI/LFM2.5-1.2B-Thinking model, produced by RACER IS OP using Heretic v1.4.0. It leverages a hybrid convolution and attention architecture with a 32,768 token context length. The primary differentiation is its significantly reduced refusal behavior, dropping from 98/100 to 10/100, achieved through targeted weight edits rather than fine-tuning, preserving the base model's reasoning capabilities.
Key Capabilities
- Explicit Reasoning Trace: Emits a
<think>...</think>trace before its final answer, providing transparency into its thought process. - Decensored Output: Refusal behavior is substantially suppressed, allowing for broader utility.
- Efficient On-Device Performance: Designed to run efficiently on devices like phones and Raspberry Pi, with quantized versions under 1GB.
- Pythonic Tool-Calling: Supports native Pythonic tool-call format for integration with external functions.
- Multilingual Coverage: Retains the base model's multilingual capabilities.
- High Reasoning Scores: Achieves strong performance on benchmarks such as MATH-500 (87.96), GSM8K (85.60), IFEval (88.42), and BFCLv3 (56.97).
Good For
- Local Agents: Ideal for on-device AI agents where cloud round-trips are a bottleneck.
- Mathematical & Multi-step Reasoning: Excels in tasks requiring complex calculations and logical steps.
- RAG Applications: Suitable for Retrieval Augmented Generation scenarios.
- Structured Extraction: Effective for extracting structured information.
- Developers Needing Unfiltered Responses: For use cases where the base model's refusal behavior is undesirable, with awareness of responsible deployment.