LFM 2.5

LFM2.5-1.2B-Instruct is a compact, instruction-tuned language model from Liquid AI, purpose-built for on-device deployment. Part of the LFM2.5 family, it delivers surprisingly strong performance from just 1.17 billion parameters — rivaling much larger models while running under 1GB of memory.

Architecture & Training

The model uses a hybrid architecture combining 10 double-gated LIV convolution blocks with 6 grouped-query attention (GQA) blocks across 16 layers. It supports a 32,768-token context window and a vocabulary of 65,536 tokens. Pre-training was extended from 10T to 28T tokens, followed by large-scale multi-stage reinforcement learning. The knowledge cutoff is mid-2024.

Capabilities

LFM2.5-1.2B-Instruct excels at agentic tasks, data extraction, and RAG workflows. It supports eight languages — English, Arabic, Chinese, French, German, Japanese, Korean, and Spanish — and includes built-in function calling with Pythonic tool-use syntax. The model is best suited for structured tasks rather than knowledge-intensive queries or programming.

Edge-First Performance

Speed is a defining feature. The model achieves 239 tok/s decode on AMD CPU and 82 tok/s on mobile NPU, with day-one support for llama.cpp, MLX, vLLM, and Transformers. Quantized variants (GGUF, ONNX, MLX) are available for optimized deployment across cloud, desktop, and mobile environments.

Benchmarks at a Glance

Among sub-2B models, LFM2.5-1.2B-Instruct leads on GPQA (38.89), MMLU-Pro (44.35), IFEval (86.23), and AIME25 (14.00), outperforming Qwen3-1.7B, Granite 4.0-1B, Llama 3.2-1B, and Gemma 3-1B across most benchmarks.

Fine-Tuning

The model supports fine-tuning via Unsloth and TRL, including SFT, DPO, GRPO, and continued pre-training — making it highly adaptable for domain-specific use cases.