sonic-coder/LFM2.5-350M-heretic

Overview

sonic-coder/LFM2.5-350M-heretic is a 0.35 billion parameter instruction-tuned language model, derived from LiquidAI's LFM2.5-350M. Its primary distinction is being a decensored version, achieved through the application of the Heretic v1.1.0 tool. This modification significantly reduces the model's refusal rate, from 88/100 in the original to 5/100 in this version, as demonstrated by comparative examples.

Key Capabilities & Features

• Decensored Output: Provides less restrictive responses compared to the base model, with a substantially lower refusal rate.
• On-Device Optimization: Designed for efficient deployment on edge devices, offering fast inference speeds (e.g., 313 tok/s on AMD CPU, 188 tok/s on Snapdragon Gen4) and low memory footprint (under 1GB).
• Extended Training: Built on the LFM2 architecture with extended pre-training (28T tokens) and large-scale multi-stage reinforcement learning.
• Multilingual Support: Supports English, Arabic, Chinese, French, German, Japanese, Korean, Portuguese, and Spanish.
• Tool Use: Capable of function calling, supporting both Pythonic and JSON function calls, and interpreting tool outcomes.
• Context Length: Features a substantial context window of 32,768 tokens.

Performance & Benchmarks

While the model's core strength lies in its decensored nature, it also demonstrates competitive performance against other small models (e.g., LFM2-350M, Granite 4.0-350M) across various benchmarks like GPQA Diamond, IFEval, and Multi-IF, showing improvements over its predecessor, LFM2-350M.

Good for

• Applications requiring less restrictive content generation.
• Data extraction and structured outputs.
• Tool use and function calling scenarios.
• On-device and edge deployments where computational resources are limited.

Not Recommended for

• Knowledge-intensive tasks.
• Programming tasks.