Liquid AI Shares Recipe for 1B On-Device Phone Agents
maximelabonne · x · 2026-08-03
Liquid AI's head of post-training detailed the full tech stack for building high-performance Small Language Models (SLMs). The presentation explains how to train a model under 1GB that runs efficiently on-device in 20 minutes.
The core training stack includes:
- LFM2.5 as the base model
- On-policy preference alignment (DPO)
- Agentic reinforcement learning
- Curriculum training combined with iterative model merging
Through this loop, frontier small models can now outperform traditional 70B models on critical agentic tasks like tool calling.
Related event: Liquid AI Unveils Training Stack for 1B On-Device Models(2 posts)→
More from Models
- DeepSeek V4F Beats V4P at Chess, Writes Its Own Engine Mid-Game — teortaxesTex · 2026-08-03
- Claude Opus Falsely Claims Job Done When Context Window Fills Up — pvncher · 2026-08-03
- DeepSeek Processes 8T Tokens Daily, MiniMax Open-Sources Video Model H3 — 快鲤鱼 · 2026-08-03
- SenseNova Open-Sources 8B Multimodal Model U1.5-Lite: Native 4K Generation & Precise Editing — 量子位 · 2026-08-03
- Migrating from GPT-4o to GPT-5.1: Handling RAG Agent Response Style Regressions — IncreaseLocal2574 · 2026-08-03
- Alibaba Releases Most Capable AI Model Qwen3.8-Max, Challenging OpenAI and Anthropic — The Verge AI · 2026-08-03