LFM2.5-2.6B: A New Small Language Model for On-Device Agents
Liquid AI
Liquid AI has released LFM2.5-2.6B, a compact 2.6-billion-parameter model optimized for agentic tasks with efficient on-device deployment. It rivals models up to four times larger on agentic, instruction-following and tool-use benchmarks, and offers high inference speeds on both CPU and GPU.
Liquid AI introduced LFM2.5-2.6B, a 2.6 billion parameter model designed for agentic tasks at the edge. Pre-trained on about 34 trillion tokens with a context window of 128K, it underwent post-training in four stages: supervised fine-tuning, teacher specialization, multi-domain on-policy distillation, and agentic reinforcement learning within real agent harnesses. Benchmarking shows it competes with notably larger models, excelling in instruction following and tool use, while larger models maintain a lead in coding tasks. The model achieves inference speeds of 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen CPU, and can even run on a phone at 30 tokens per second. Support is available day-one in llama.cpp, MLX, vLLM, SGLang, and ONNX. LFM2.5-2.6B and its base variant are available on Hugging Face.
- Abbreviations
- SFT = Supervised Fine-Tuning — обучение с учителем
- MOPD = Multi-domain On-Policy Distillation — многодоменная дистилляция на политике
- RL = Reinforcement Learning — обучение с подкреплением
- CPU = Central Processing Unit — центральный процессор
- GPU = Graphics Processing Unit — графический процессор
- ONNX = Open Neural Network Exchange — открытый формат обмена нейронными сетями
- SGLang = Structured Generation Language — язык структурированной генерации
- MLX = Machine Learning Framework for Apple Silicon — фреймворк машинного обучения для Apple Silicon
- vLLM = Virtual Large Language Model — виртуальная большая языковая модель
Source: Hugging Face blog —
original