Skip to content

Maxime Labonne

Liquid AI

  • 1 episode on Chain of Thought
  • 6 topics covered
  • 1 full transcript
  • Nov 2025 first on the show
Maxime Labonne

Maxime Labonne is Head of Post-Training at Liquid AI and the creator of the widely starred LLM Course on GitHub. On Chain of Thought he challenged the transformer monopoly, explaining how Liquid AI’s hybrid architecture delivers faster inference and a smaller footprint so models can run on phones and laptops.

Cover art for Beyond Transformers: How Liquid AI Is Rethinking LLM Architecture | Maxime Labonne

Latest episode with Maxime Labonne · EP 43

Beyond Transformers: How Liquid AI Is Rethinking LLM Architecture | Maxime Labonne

Maxime Labonne, Liquid AI ·

Show notes & transcript → All listening options →

Episodes

What the conversations cover

These notes summarize the episodes featuring Maxime Labonne, including other speakers in panel conversations. Follow a source link to read the discussion in context.

EP 43 Beyond Transformers: How Liquid AI Is Rethinking LLM Architecture | Maxime Labonne

  • Liquid AI’s LFM2 models are a hybrid architecture: each model carries six attention layers and ten short convolution layers — more convolution than attention. The convolution layers keep the KV cache from exploding on long context, so inference is much faster and memory usage much lower while matching a pure transformer’s quality. Episode takeaways
  • Liquid open-sourced 17 models in roughly three months (since July). The LFM2 line ships at 350M, 700M, and 1.2B parameters, designed first for phones and wearables rather than scaled up — the opposite of the “how big can we make this damn thing” race. Episode takeaways
  • Theoretical speed gains might not be realized on real hardware. After LFM1, Liquid started optimizing LFM2 directly on a target device (a Samsung phone) plus over 100 pre-training evaluations, so an operator that looks fast on paper is actually measured fast before it ships. Episode takeaways

In their words

AI and machine learning going forward? Yeah. So open source for me is quite selfish, honestly. It's mostly me trying to learn stuff online, and then it gets picked up a lot. Mhmm. But I think it's actually quite important to do it for yourself and not for other people because, well, this is your own interest. Right? And this is an incredible way of learning new things.

Maxime Labonne Beyond Transformers: How Liquid AI Is Rethinking LLM Architecture | Maxime Labonne · 42:15

Transcript excerpts; ellipses mark omitted words. Some automatically selected quotes have filler words removed.