One of the changes is in the Michi component, the part whose job is detecting the language a user is using before their message gets passed on to Siesta. Its scope was adjusted — from the original plan of supporting dozens of languages, it's now focused on 14 languages — and pretraining itself is now complete, after its status was still empty in the previous version.
Meanwhile, development of a far larger-scale model — the v6 generation with over 4 billion parameters — is still ongoing, and the process hasn't been without friction. There are still a few technical hurdles around context-size configuration and how to split the computational load across multiple TPU chips at once that need resolving first.
Preprint v1.1 — September 2026
The full manuscript, including technical details not covered here, can be read directly on Zenodo.
Read the Full Paper →For anyone not yet familiar with the overall background of Siesta's research, there's a previous article that covers the story from the start.