Author(s): Chew Loong Nian – AI ENGINEER

Originally published on Towards AI.

I did not expect a model that activates 1.5 billion parameters to walk all over one that activates 4 billion — and to do it on a benchmark that actually matters for agents. But that is exactly what happened when I put Liquid AI’s new LFM2.5–8B-A1B (released May 28, 2026) up against Google’s Gemma 4 26B. On Tau²-Telecom, a hard multi-turn tool-use benchmark, the tiny Liquid model scored 88.07 against Gemma 4 26B’s 42.11. That is a 46-point gap — in favor of the model small enough to run on my laptop with memory to spare.

After the lead, the article explains why an 8B MoE model can matter again for on-device agents: it’s optimized for privacy-sensitive, multi-step tool use and reliable abstention, not just leaderboard-style generality. It details the LFM2.5–8B-A1B architecture (sparse MoE with many short-range gated convolution layers instead of attention) and notes a reasoning-only design intended to improve tool/agent behavior without making inference too expensive on edge hardware. The author walks through benchmarks showing major gains over the prior LFM2 generation and a large Tau²-Telecom advantage over Gemma 4 26B, including a big improvement in non-hallucination/hedging. They then cover practical deployment (vLLM, Transformers, and llama.cpp/quantized GGUF on CPU), runtime performance and tool-call formatting, and the “catch” that the model is strong for agentic tool dispatch but weak for coding and some knowledge-intensive tasks. The piece concludes with guidance on when to choose this model (private on-device tool-calling) versus when to use something else (dedicated coding models or systems with retrieval for broad knowledge), culminating in a brief verdict that Liquid built a “scalpel” for agents rather than a “Swiss Army knife.”

Read the full blog for free on Medium.

Published via Towards AI