Dingdust/LFM2.5-1.2B-Thinking-heretic

Model Overview

Dingdust/LFM2.5-1.2B-Thinking-heretic is a 1.2 billion parameter language model, a decensored variant of Liquid AI's LFM2.5-1.2B-Thinking, built with Heretic v1.4.0. It is part of the LFM2.5 family, optimized for on-device deployment with extended pre-training (28T tokens) and multi-stage reinforcement learning. The model boasts a 32,768 token context length and supports English, Arabic, Chinese, French, German, Japanese, Korean, and Spanish.

Key Capabilities

• On-Device Performance: Rivals larger models in performance while running under 1GB of memory, with fast inference speeds (e.g., 239 tok/s decode on AMD CPU, 82 tok/s on mobile NPU). It has day-one support for llama.cpp, MLX, and vLLM.
• Reasoning Focus: Specifically designed as a general-purpose reasoning model, distinguishing it from instruction-tuned or base variants.
• Tool Use: Supports function calling with a flexible mechanism for defining tools in the system prompt and interpreting function execution results.
• Decensored Output: Achieves significantly lower refusal rates (10/100) compared to the original model (97/100), indicating a less restrictive output policy.
• Long-Context Scalability: Demonstrates robust long-context inference, sustaining high decoding throughput even at its full 32K context length on optimized hardware.

When to Use This Model

• Agentic Tasks: Ideal for applications requiring autonomous decision-making and interaction.
• Data Extraction: Well-suited for extracting structured information from unstructured text.
• Retrieval-Augmented Generation (RAG): Effective for scenarios where external knowledge retrieval enhances generation quality.
• Edge Deployment: Excellent choice for applications requiring high performance on resource-constrained devices like mobile phones, IoT, and embedded systems.

Limitations

• Not Recommended for: Knowledge-intensive tasks and programming, as explicitly stated by the developers.