Frontier Agentic Models for the Edge
Session Overview
A new class of small agentic models is emerging β models that can call tools and complete tasks while running entirely on phones and laptops. This session breaks down how to post-train these models using the LFM2.5 recipe: on-policy preference alignment, agentic reinforcement learning, and curriculum training with iterative model merging.
Maxime will cover the training challenges unique to the 1B scale β doom loops, capability interference, and how to fix them β and share a concrete playbook for fine-tuning and deploying small models across real use cases, from structured data extraction to multi-turn tool use.
What you'll take away
- The LFM2.5 post-training recipe β on-policy preference alignment, agentic reinforcement learning, and curriculum training with iterative model merging.
- Solving 1B-scale training challenges β how to diagnose and fix doom loops and capability interference.
- A deployable playbook β fine-tune and ship small agentic models for tasks from structured data extraction to multi-turn tool use.
The goal is to give practitioners a practical, reproducible playbook for building capable agentic models small enough to run on the edge.
Speaker
Maxime Labonne is Head of Post-Training at Liquid AI. He holds a Ph.D. in Machine Learning from the Polytechnic Institute of Paris and is a Google Developer Expert in AI/ML.
He has made significant contributions to the open-source community, including best-in-class models like LFM2/2.5 and the popular LLM Course. He is the author of the best-selling books "LLM Engineer's Handbook" and "Hands-On Graph Neural Networks Using Python."