Liquid AI Releases LFM2.5-2.6B: Efficient On-Device Agentic Model
Liquid AI has introduced LFM2.5-2.6B, a model boasting 2.6 billion parameters designed for on-device agentic AI tasks. This model has been pre-trained on an impressive 34 trillion tokens and includes a context window of 128K tokens. Its post-training process features supervised fine-tuning, teacher specialization, multi-domain on-policy distillation, and agentic reinforcement learning. The Agentic RL pipeline enhances performance through a Sandbox Service and Blackbox Harness, utilized by agents such as OpenClaw. Competing with larger models, LFM2.5-2.6B excels in instruction adherence and tool usage, achieving speeds of 220 tokens per second on the Apple M5 Max and 113 on the AMD Ryzen AI Max+ 395, while maintaining under 2.5 GB of memory usage. It is compatible with llama.cpp, MLX, vLLM, SGLang, and ONNX, and can be accessed on Hugging Face.
Key facts
- LFM2.5-2.6B is a 2.6B parameter model for on-device agentic AI.
- Pre-trained on ~34T tokens with 128K context window.
- Post-training: SFT, teacher specialization, MOPD, Agentic RL.
- Agentic RL uses Sandbox Service, Blackbox Harness, Harness Proxy.
- Benchmarks: competes with models 4x larger, tops instruction following and tool use.
- Inference: 220 tok/s on Apple M5 Max, 113 tok/s on AMD Ryzen CPU.
- Supports llama.cpp, MLX, vLLM, SGLang, ONNX.
- Available on Hugging Face today.
- Browser demo and guides for OpenClaw, Hermes Agent, Pi.
Entities
Institutions
- Liquid AI
- Hugging Face