unsloth/LFM2-350M

LFM2-350M: A Hybrid Model for Edge AI

LFM2-350M, developed by Liquid AI, is a 0.35 billion parameter hybrid model specifically engineered for efficient edge AI and on-device deployment. It integrates a novel architecture combining multiplicative gates and short convolutions, featuring 10 double-gated short-range LIV convolution blocks and 6 grouped query attention blocks. This design contributes to its superior performance in speed and memory efficiency.

Key Capabilities & Features

• Optimized Performance: Achieves 3x faster training and 2x faster decode/prefill speeds on CPU compared to Qwen3, outperforming similarly-sized models across various benchmarks including knowledge, mathematics, instruction following, and multilingual tasks.
• Flexible Deployment: Designed to run efficiently on CPU, GPU, and NPU hardware, making it suitable for deployment on smartphones, laptops, and vehicles.
• Multilingual Support: Supports English, Arabic, Chinese, French, German, Japanese, Korean, and Spanish.
• Tool Use: Incorporates a structured tool-use mechanism, allowing for function definition, calling, execution, and interpretation within conversations.
• Training: Utilizes knowledge distillation from LFM1-7B, large-scale SFT, custom DPO with length normalization, and iterative model merging, trained on 10 trillion tokens.

Recommended Use Cases

LFM2-350M is particularly suited for fine-tuning on narrow use cases to maximize performance. It is recommended for:

• Agentic tasks
• Data extraction
• RAG (Retrieval Augmented Generation)
• Creative writing
• Multi-turn conversations

However, it is not recommended for knowledge-intensive tasks or those requiring programming skills due to its small size.