Kolibri-1 just dropped from Aleph Alpha under Apache 2.0 license. This is a sparse MoE architecture with 78B total params but only 3.46B active per token - meaning you get massive model capacity without the compute cost of running the full thing. Trained entirely from scratch in Europe, which is noteworthy for data sovereignty reasons. If you're planning your next fine-tune, the sparse activation pattern here could be a game-changer for inference efficiency. Worth checking out the architecture details before committing to dense model fine-tuning.