Maxime Labonne is Head of Post-Training at Liquid AI and creator of the LLM Course, an open-source guide to developing and deploying language models. He builds small foundation models designed for the constraints of actual devices while making advanced training techniques accessible to other engineers.
From cybersecurity to foundation models
Labonne earned a doctorate in machine learning at the Institut Polytechnique de Paris, researching anomaly-based network intrusion detection with the French Alternative Energies and Atomic Energy Commission. He subsequently held machine-learning research positions at Airbus and JPMorgan Chase before joining Liquid AI.
His earlier work included graph neural networks and the book Hands-On Graph Neural Networks Using Python. He later created the LLM Course, combining learning roadmaps with practical notebooks, and co-wrote LLM Engineer’s Handbook with Paul Iusztin. His other open-source projects include llm-datasets, a curated collection of post-training datasets and tools, and LLM AutoEval, which automates model evaluation in Google Colab.
- Evaluate before fine-tuning. Labonne recommends defining quality, cost, and latency metrics, then testing prompting before committing to customization. Effective post-training depends on accurate, diverse, challenging examples; deduplication, reward models, and model-based judging improve dataset quality. His preference-optimization tutorial demonstrates how these decisions translate into practice.
- Model merging as open-source experimentation. His MergeKit work combines existing models through interpolation, parameter pruning, and layer recombination without repeating full training. His Beyonder experiment assembled specialized chat, code, role-playing, and mathematics models into a mixture of experts, initializing its router with representative prompts. He cautions that benchmark contamination can inflate apparent leaderboard gains.
- Small models require hardware-aware design. Working on Liquid Foundation Models, Labonne emphasizes that oversized embedding layers can consume parameters needed for reasoning. LFM2 combines gated short convolutions with grouped-query attention, with architectural choices tested on target CPUs and smartphones. His account of small-model development prioritizes specific capabilities such as extraction and tool use over uniform competence across every benchmark.
- Train against failure, then extend with tools. Small reasoning models can enter a doom loop, repeating text without producing an answer. Labonne addresses the problem with preference data that rejects looping responses, reinforcement learning with verifiable rewards, and repetition penalties. For on-device language models, he favors web search and Python tools to compensate for limited internal knowledge and long-context weaknesses. A browser-based vision-language demonstration reflects his interest in capable models that operate close to their users.
