Trained from scratch.
Ask it anything.

A 1.32B-parameter Mixture-of-Experts model pretrained on 18B tokens, then tuned with SFT and DPO. Only 0.28B parameters run per token. It writes and rewrites well; for anything factual, upload a document and it will answer from that.