Recurrent-SmolLM2-360M-4-14-4 (untrained scaffold)

This is not a usable model. It is SmolLM2-360M converted into a Huginn-style recurrent-depth architecture (prelude/core/coda layer split: 4/14/4), with the recurrent core layers randomly initialized. No continued-pretraining has been run on it yet.

It exists only as the starting checkpoint for a training pipeline (retrofitting-recurrence), and is public solely so that pipeline can pull it without requiring an HF token. Running inference on it as-is will produce garbage output from the randomly-initialized core layers.

If you're looking for a working model, use the base HuggingFaceTB/SmolLM2-360M instead, or check back here for a trained release under a different repo name.