LFM2.5-2.6B is a compact reasoning model from Liquid AI designed to sit inside lightweight agent and retrieval pipelines. Liquid positions it for agent workflows, data extraction, retrieval-augmented generation, and long-context processing, while steering users away from agentic coding or knowledge-heavy tasks where a larger model would be safer. Because the weights are openly published on Hugging Face under the LiquidAI namespace, teams that want full control can self-host the same checkpoint instead of relying on a hosted endpoint, and the model is offered free of charge on OpenRouter through a Liquid-hosted provider with reported round-trip latency near two seconds and throughput above two hundred tokens per second.
In practice the model fits well as a fast, low-cost reasoning layer for routing, classification, structured extraction, and tool-calling agents where the prompt and retrieved context stay within its tens-of-thousands-token working window. Its always-on reasoning signal, combined with text-in and text-out behavior and support for tool calls and structured output, makes it a flexible primitive for orchestrators that want a small, open model to handle routine decisions while reserving heavier models for complex reasoning or code generation. Teams picking this checkpoint through the Kilo Gateway get the same Liquid AI model without inference markup, which keeps experimentation costs negligible while still allowing migration to larger Liquid models if a workload outgrows the 2.6B parameter footprint.
