Liquid AI’s latest release is built for the edge, not the data center. LFM2.5-2.6B, unveiled in early August, packs agentic behavior into 2.69 billion parameters so phones, laptops, PCs, and robots can plan, call tools, and finish multi-step tasks without a network round trip.
Pre-training consumed about 34 trillion tokens, and the design pairs a 128,000-token vocabulary with a context window of 131,072 tokens. Keeping inference on-device means prompts never leave the machine and each run costs almost nothing. Liquid AI says its tool-use and instruction-following results hold up against models nearly four times its size, leading every instruction-following benchmark it reports against gemma-4 and Qwen3.5 variants, with larger models still ahead in coding.
Both checkpoints are public on Hugging Face under the lfm1.0 license, with native, GGUF, MLX, and ONNX weights and day-one support in llama.cpp, vLLM, SGLang, and LM Studio. Performance claims include 220 tokens per second on an M5 Max in under 2.5 GB of memory, and a single H100 handling about 1.3 billion tokens per day.
Getting there took a four-stage post-training recipe: two supervised fine-tuning rounds, domain-specialist teachers trained with verifiable rewards, multi-domain on-policy distillation, and agentic reinforcement learning with GRPO inside real harnesses like Hermes Agent and OpenClaw.
Liquid AI is targeting automotive, consumer electronics, industrial robotics, healthcare, finance, e-commerce, and defense, suggesting the model for on-device assistants, offline document triage, form extraction, and robotics command parsing rather than agentic coding.
