MII-LLM releases Zagreus, Nesso small models

This is a strong example of small-model engineering done seriously: the value is not just the weights, but the full reproducible pipeline from tokenization to Slurm orchestration to post-training. In the sub-1B regime, disciplined data and training choices matter more than architectural novelty.

• –Dense 0.4B is the sensible call here; MoE complexity is hard to justify when stability and hardware utilization are the bottlenecks.
• –The bilingual English + target-language setup is a practical way to cover European languages without pretending a tiny model can be universal.
• –Nesso-agentic is likely the most useful checkpoint in the set because structured output and function calling are where small models can still feel “product-ready.”
• –The benchmark story is encouraging, but the real ceiling remains visible: arithmetic, factual recall, and repetition are still weak points.
• –The open variant is the most interesting piece for the ecosystem, because reproducible small-model training is still rare and highly transferable.