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

TEXT GENERATIONConcurrent Unit Cost:1Model Size:1.2BQuant:BF16Context Size:32kPublished:Aug 9, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

Dingdust/LFM2.5-1.2B-Instruct-heretic is a 1.2 billion parameter instruction-tuned causal language model, based on the Liquid AI LFM2.5 architecture, with a 32,768 token context length. This model is a decensored version, created using Heretic v1.4.0, and is optimized for on-device deployment with fast edge inference. It offers best-in-class performance among sub-2B models, excelling in agentic tasks, data extraction, and RAG.

Loading preview...

Model Overview

Dingdust/LFM2.5-1.2B-Instruct-heretic is a 1.2 billion parameter instruction-tuned model derived from Liquid AI's LFM2.5 architecture, specifically modified using Heretic v1.4.0 for decensored responses. It features a substantial 32,768 token context length and was trained on an extended 28 trillion tokens. The model is designed for efficient on-device deployment, demonstrating fast inference speeds on various hardware, including CPUs and NPUs, while maintaining a low memory footprint.

Key Capabilities & Features

  • Decensored Output: Modified to provide less restrictive responses compared to its original counterpart, with significantly reduced refusals (3/100 vs. 98/100).
  • On-Device Optimization: Engineered for fast edge inference, achieving 239 tok/s decode on AMD CPU and 82 tok/s on mobile NPU, running under 1GB of memory.
  • Multilingual Support: Supports English, Arabic, Chinese, French, German, Japanese, Korean, and Spanish.
  • Tool Use: Integrates function calling capabilities, allowing for structured interaction with external tools via Pythonic or JSON function calls.
  • Strong Performance: Outperforms other sub-2B models like Qwen3-1.7B and Llama 3.2 1B Instruct across several benchmarks including GPQA, MMLU-Pro, IFEval, and BFCLv3.

Recommended Use Cases

  • Agentic Tasks: Ideal for applications requiring autonomous decision-making and interaction.
  • Data Extraction: Effective for extracting specific information from text.
  • Retrieval Augmented Generation (RAG): Suitable for enhancing generation with retrieved knowledge.

Limitations

  • Not recommended for knowledge-intensive tasks or programming-specific applications.