I used Claude to build a 135M looped LLM from scratch — it debugged Parcae paper implementations, caught optimizer routing bugs, wrote Modal training infra, and helped ship to HuggingFace. Here's t...

Built a 135M dense looped LLM from scratch. Spent 2 weeks debugging Parcae's LTI stability mechanisms across 5 ablations. None of them beat the naive baseline at this scale. Trained for real anyway. SFT'd it. Shipped it. Here's the full honest story. What I built A 135M parameter looped transformer trained from scratch on FineWeb (4.6B tokens), inspired by the Parcae paper (arXiv:2604.12946 — "Scaling Laws For Stable Looped Language Models"). 🤗 Base model: huggingface.co/harims95/LoopLM-135M-naive 🤗 SFT mod