Liquid AI has released LFM2.5-230M, a 230-million-parameter model that outperforms models four times its size at data extraction and structured tool calls. The release challenges the prevailing assumption that bigger models are always better for agentic workflows.
Smaller, But Sharper
While the AI industry has raced toward larger models — 3-billion-parameter models like VibeThinker solving advanced calculus — Liquid AI took a different path. LFM2.5-230M is optimized specifically for structured tool calls and keeping agentic pipelines running efficiently.
The key insight: not every task requires a massive model. For structured outputs like API calls, tool invocations, and data extraction, a smaller model trained on the right data can outperform giants.
Designed to Run Anywhere
Perhaps most significantly, LFM2.5-230M is designed to run “anywhere” — on edge devices, in browsers, on mobile phones. This positions Liquid AI’s approach for applications where latency, cost, or infrastructure constraints make larger models impractical.
For enterprises, this means agentic pipelines can be powered by lightweight models for routine tasks, reserving larger models for genuinely complex reasoning — a more economically rational architecture.
The Efficiency Argument
The broader context is economic pressure on AI deployments. As organizations scale AI agents across workflows, the cost of inference becomes a significant line item. LFM2.5-230M suggests that thoughtful architecture — matching model size to task complexity — can deliver both better economics and better performance.
This aligns with a growing recognition in the industry that the “brute force” approach to scaling model size has diminishing returns for many practical applications. Liquid AI’s work represents a counter-narrative worth watching.