This feels like the first open model that doesn’t require significant caveats when comparing to frontier proprietary models. The parameter efficiency alone suggests some genuine innovations in training methodology. I am keen to see some independent verification of the results and to see how if does on Aider’s LLM Leaderboard.