Liquid AI released LFM2.5-2.6B, an open-weight, 2.6-billion-parameter model built specifically to run agentic workloads — planning, tool calling, multi-step task execution — entirely on local hardware rather than through a cloud API. What makes this release notable for developers isn't the parameter count itself, which is small by frontier standards, but the deliberate choice to optimize the entire training pipeline around agent behavior rather than general chat quality. The model went through a four-stage post-training recipe: broad supervised fine-tuning, a distillation step where domain-specialist teacher models feed a shared student model token-level guidance, and a final reinforcement learning stage run inside real agent harnesses (Hermes Agent, OpenClaw and others) so the model learns the actual system prompts, tool schemas, and interaction quirks of the tools it will be deployed with, not just abstract instruction-following. The practical payoff is speed and cost, not raw capability. On a laptop-class Apple M5 Max chip it decodes around 220 tokens per second while staying under 2.5GB of memory, and it still holds roughly 30 tokens per second on a phone. Because there's no per-token API cost once the weights are downloaded, teams can run large numbers of parallel background agents around the clock without watching a usage bill climb — a genuinely different design point than routing everything through a hosted frontier model. On benchmarks for instruction-following and tool use it beats or matches models roughly four times its size, though it still trails larger models specifically on coding-heavy tasks. Base and post-trained checkpoints are both on Hugging Face with day-one support for llama.cpp, MLX, vLLM, SGLang and ONNX, so it's usable today across edge devices, laptops and GPU servers without extra conversion work. For builders who want agent behavior without every tool call round-tripping to a cloud model, this is a directly deployable option rather than a research curiosity.