MuXodious/LFM2.5-1.2B-Instruct-absolute-heresy
MuXodious/LFM2.5-1.2B-Instruct-absolute-heresy is a 1.2 billion parameter instruction-tuned model based on the LFM2.5 architecture, developed by Liquid AI. This model is a fine-tune produced via P-E-W's Heretic engine, resulting in an "Absolute Heresy" classification due to low refusals and KL Divergence. It is optimized for on-device deployment, offering best-in-class performance and fast edge inference for agentic tasks, data extraction, and RAG.
Loading preview...
LFM2.5-1.2B-Instruct-absolute-heresy Overview
This model is a 1.2 billion parameter instruction-tuned variant of Liquid AI's LFM2.5 architecture, specifically processed using P-E-W's Heretic engine. It achieves an "Absolute Heresy" classification, indicating a low refusal rate (7/100) and low KL Divergence (0.0679), suggesting a significant deviation from its base model's original doctrine while maintaining coherence. The LFM2.5 family is designed for on-device deployment, offering performance comparable to much larger models.
Key Capabilities & Features
- Optimized for Edge Inference: Achieves 239 tok/s decode on AMD CPU and 82 tok/s on mobile NPU, running under 1GB of memory.
- Broad Compatibility: Day-one support for
llama.cpp, MLX, andvLLMfor diverse deployment scenarios. - Scaled Training: Benefits from extended pre-training on 28 trillion tokens and large-scale multi-stage reinforcement learning.
- Multilingual Support: Handles English, Arabic, Chinese, French, German, Japanese, Korean, and Spanish.
- Tool Use: Supports function calling with a Pythonic format, enabling agentic tasks.
- High Context Length: Features a 32,768-token context window.
Performance & Benchmarks
LFM2.5-1.2B-Instruct demonstrates strong performance against other sub-2B models across various benchmarks, including GPQA (38.89), MMLU-Pro (44.35), and IFEval (86.23), often outperforming models like Qwen3-1.7B and Gemma 3 1B IT in its class.
Recommended Use Cases
- Agentic Tasks: Ideal for applications requiring structured interactions and decision-making.
- Data Extraction: Effective for extracting specific information from text.
- Retrieval Augmented Generation (RAG): Suitable for enhancing generation with external knowledge.
Limitations
- Not Recommended for Knowledge-Intensive Tasks: May not excel in areas requiring deep factual recall.
- Not Recommended for Programming: Performance in code generation or understanding is not a primary strength.