Fine-tune LiquidAI/LFM2.5-1.2B-Instruct with Unsloth and Hugging Face Jobs (free credits)

AI Impact Summary

The post describes using Unsloth with Hugging Face Jobs to fine-tune LiquidAI/LFM2.5-1.2B-Instruct via coding agents (Claude Code, Codex) using the hf jobs CLI. Unsloth claims ~2x training speed and ~60% VRAM savings, enabling small-model fine-tuning at a fraction of typical cloud GPU costs and even on-device deployment potential. Operational considerations include free credits and required HF billing setup, plus monitoring via Trackio and credential management (HF_TOKEN, write permissions) to manage costs and job status.