Alibaba drops Marco-Mini, Marco-Nano models

This is more interesting as an efficiency signal than a raw-size story. Alibaba is pushing sparse MoE into a practical open-weights release, and the active-parameter ratios are the headline.

• –Marco-Mini uses 256 experts with 8 active per token; Marco-Nano uses 232 experts with 8 active, so the design is tuned for low compute per token.
• –Both models are Apache 2.0 on Hugging Face and cover 29 languages, which makes them immediately relevant for local deployment and multilingual apps.
• –The releases are positioned against Qwen3, Gemma3, Ministral3, Granite4, and LFM2, so Alibaba is clearly targeting the small-to-mid instruct model tier rather than chasing giant parameter counts.
• –The benchmark numbers look strong, but the real differentiator will be latency, memory use, and whether the quality holds up outside curated evals.
• –The post-training recipe matters too: both models are upcycled from Qwen3-0.6B-Base and then refined with SFT plus distillation, which is a very Alibaba-style efficiency play.