LFM 2

LFM2-1.2B is a 1.2-billion-parameter hybrid language model from Liquid AI, purpose-built for edge AI and on-device deployment. Part of the second-generation LFM family, it combines a novel architecture of multiplicative gates and short convolutions — specifically 10 double-gated short-range LIV convolution blocks and 6 grouped query attention (GQA) blocks — to deliver strong quality at minimal resource cost.

Key Strengths

• Speed — 3× faster training than the previous LFM generation; 2× faster CPU decode and prefill compared to Qwen3.
• Quality — Outperforms similarly-sized models across knowledge, math, instruction following, and multilingual benchmarks.
• Flexible deployment — Runs efficiently on CPU, GPU, and NPU hardware, targeting smartphones, laptops, and vehicles.

The model supports a 32,768-token context window and eight languages: English, Arabic, Chinese, French, German, Japanese, Korean, and Spanish. It was trained on 10 trillion tokens (≈75% English, 20% multilingual, 5% code) using knowledge distillation from LFM1-7B, large-scale supervised fine-tuning, custom DPO, and iterative model merging.

Benchmarks

LFM2-1.2B leads or closely matches larger competitors on automated evaluations including MMLU, GPQA, IFEval, and MGSM.

LLM-as-a-Judge evaluations confirm strong conversational and instruction-following quality relative to its size class.

Inference Performance

On-device throughput tests show clear advantages on CPU runtimes like ExecuTorch and llama.cpp.

Recommended Use Cases

Due to its compact size, Liquid AI recommends fine-tuning LFM2-1.2B on focused tasks for best results. It excels at agentic workflows, data extraction, RAG, creative writing, and multi-turn conversations. Built-in tool-use support enables structured function calling via a ChatML-style template. Compatible with Hugging Face Transformers (v4.55+), vLLM, and llama.cpp via GGUF.