Liquid AI releases LFM2.5-2.6B, a 2.6B-parameter on-device agentic model with tool-calling and multi-step workflows

Liquid AI published LFM2.5-2.6B, a 2.6B-parameter model optimized for on-device agentic use cases such as tool calling and multi-step workflows. The model is intended to run on everyday hardware, including laptops and phones, enabling developers to deploy agents locally without sending data to the cloud.

The model achieves competitive performance on agentic benchmarks while remaining small and fast. Liquid AI reports inference speeds of 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen AI Max+ 395 CPU, with memory usage under 2.5 GB. On GPU, the model reaches nearly 15,000 output tokens per second at high concurrency, enabling high-volume workloads on a single H100.

LFM2.5-2.6B was trained using a multi-stage post-training process that includes supervised fine-tuning focused on agentic data, teacher specialization across domains, multi-domain on-policy distillation, and agentic reinforcement learning within real agent harnesses. The training pipeline separates model optimization, inference, and environment execution into distinct components: a Training Engine, a Rollout Engine, and an RL framework that orchestrates the training loop.

In benchmark comparisons against models up to roughly four times its size, LFM2.5-2.6B ranked highest on instruction-following and tool-use tasks, and performed comparably on agentic tasks. It led on knowledge benchmarks and remained close on math, though larger models retained an advantage in coding tasks.

The model ships with day-one support across multiple inference backends, including llama.cpp, MLX, vLLM, SGLang, and ONNX. Developers can access LFM2.5-2.6B and its base variant on Hugging Face, with a browser-based WebGPU demo available for testing agentic workflows without local setup.