tttt111/test-tmp3

TEXT GENERATIONConcurrent Unit Cost:1Model Size:1.2BQuant:BF16Context Size:32kPublished:Jul 2, 2026License:lfm1.0Architecture:Transformer Featherless Exclusive Cold

LFM2.5-1.2B-Instruct by Liquid AI is a 1.17 billion parameter instruction-tuned hybrid model from the LFM2.5 family, featuring a 32,768 token context length. It is specifically designed for efficient on-device deployment, offering fast inference speeds on various hardware including CPUs and NPUs. This model excels in agentic tasks, data extraction, and RAG, supporting tool use and multilingual capabilities across 8 languages.

Loading preview...

LFM2.5-1.2B-Instruct: On-Device AI

LFM2.5-1.2B-Instruct, developed by Liquid AI, is a 1.17 billion parameter instruction-tuned model built on the LFM2 architecture. It features an extended pre-training of 28 trillion tokens and large-scale multi-stage reinforcement learning. This model is optimized for on-device deployment, delivering high performance comparable to much larger models while maintaining a low memory footprint (under 1GB).

Key Capabilities

  • Efficient On-Device Inference: Achieves 239 tok/s on AMD CPU and 82 tok/s on mobile NPU, with day-one support for llama.cpp, MLX, and vLLM. Optimized versions are available for Apple Silicon and various hardware.
  • Long Context Window: Supports a context length of 32,768 tokens.
  • Multilingual Support: Capable in English, Arabic, Chinese, French, German, Japanese, Korean, and Spanish.
  • Tool Use: Supports function calling with a Pythonic format for agentic workflows, allowing for function definition, execution, and interpretation.
  • Strong Performance: Benchmarks show competitive results against other sub-2B models, particularly in IFEval (86.23) and Multi-IF (60.98), and significantly faster inference speeds compared to Qwen3-1.7B on mobile CPUs.

Good For

  • Agentic tasks
  • Data extraction
  • Retrieval Augmented Generation (RAG)
  • Edge deployment scenarios on mobile devices, laptops, and embedded systems.