unsloth/LFM2-1.2B
LFM2-1.2B is a 1.2 billion parameter hybrid model developed by Liquid AI, featuring a novel architecture with multiplicative gates and short convolutions. Designed for edge AI and on-device deployment, it offers significantly faster training and inference speeds compared to previous generations and other similarly sized models. This model excels across various benchmarks including knowledge, mathematics, instruction following, and multilingual capabilities, making it ideal for agentic tasks, data extraction, RAG, and creative writing on resource-constrained hardware.
Loading preview...
LFM2-1.2B: A Hybrid Model for Edge AI
LFM2-1.2B, developed by Liquid AI, is a 1.2 billion parameter hybrid model specifically engineered for efficient edge AI and on-device deployment. It introduces a new architecture combining multiplicative gates and short convolutions, resulting in 3x faster training and 2x faster decode/prefill speeds on CPU compared to Qwen3.
Key Capabilities & Features
- Optimized Performance: Outperforms similarly-sized models in knowledge, mathematics, instruction following, and multilingual benchmarks.
- Novel Architecture: Utilizes 10 double-gated short-range LIV convolution blocks and 6 grouped query attention (GQA) blocks.
- Flexible Deployment: Designed for efficient operation on CPU, GPU, and NPU hardware, suitable for smartphones, laptops, and vehicles.
- Multilingual Support: Supports English, Arabic, Chinese, French, German, Japanese, Korean, and Spanish.
- Tool Use: Features a structured tool-use mechanism with JSON function definitions and Pythonic function calls.
- Training: Benefits from knowledge distillation, large-scale SFT on diverse tasks, custom DPO, and iterative model merging.
Recommended Use Cases
LFM2-1.2B is particularly suited for fine-tuning on narrow use cases to maximize performance. It is recommended for:
- Agentic tasks
- Data extraction
- Retrieval Augmented Generation (RAG)
- Creative writing
- Multi-turn conversations
However, it is not recommended for knowledge-intensive tasks or those requiring advanced programming skills. The model has a context length of 32,768 tokens and was trained on a budget of 10 trillion tokens.