LiquidAI/LFM2.5-1.2B-Thinking
LiquidAI's LFM2.5-1.2B-Thinking is a 1.17 billion parameter hybrid model designed for on-device deployment, building on the LFM2 architecture with extended pre-training on 28T tokens and large-scale multi-stage reinforcement learning. This general-purpose reasoning model excels at agentic tasks, data extraction, and RAG, offering best-in-class performance for its size. It features a 32,768-token context length and supports fast edge inference, running under 1GB of memory on various devices including CPUs and NPUs.
Loading preview...
LFM2.5-1.2B-Thinking: On-Device Reasoning
LFM2.5-1.2B-Thinking, developed by LiquidAI, is a 1.17 billion parameter model optimized for on-device deployment. It is part of the LFM2.5 family, featuring a hybrid architecture with extended pre-training on 28 trillion tokens and large-scale multi-stage reinforcement learning. This model is specifically designed for reasoning tasks, offering high performance in a compact footprint.
Key Capabilities & Features
- On-Device Performance: Delivers best-in-class performance for a 1.2B model, rivaling much larger models, with fast edge inference (e.g., 239 tok/s decode on AMD CPU, 82 tok/s on mobile NPU).
- Memory Efficiency: Operates under 1GB of memory, with day-one support for
llama.cpp, MLX, andvLLMfor broad compatibility. - Extended Context: Features a substantial 32,768-token context length, enabling robust long-context inference, sustaining high decode throughput even at full context on NPUs.
- Multilingual Support: Supports English, Arabic, Chinese, French, German, Japanese, Korean, and Spanish.
- Tool Use: Integrates function calling capabilities, allowing for agentic workflows with structured tool definitions and execution.
Benchmarks & Performance
The model demonstrates strong performance across various benchmarks, particularly in reasoning and instruction-following tasks, often outperforming other sub-2B models like Qwen3-1.7B (thinking mode) in areas like GPQA Diamond, IFEval, and MATH-500. It also shows superior inference speed and lower memory usage compared to similar-sized models, making it ideal for resource-constrained environments.
Recommended Use Cases
- Agentic Tasks: Well-suited for applications requiring autonomous decision-making and task execution.
- Data Extraction: Effective for extracting structured information from text.
- Retrieval Augmented Generation (RAG): Can be integrated into RAG systems for enhanced knowledge retrieval and generation.
It is not recommended for knowledge-intensive tasks or programming.