Removing the decay phase turns training length into a flexible decision rather than a fixed plan, which matters for labs managing compute budgets mid-run. Teams planning large pre-training runs can weigh checkpoint averaging against cosine and WSD schedules using this compact comparison of named labs' choices. It is also a useful reference point when evaluating which labs disclose training details at all.
#Kolibri
德国公司 Aleph Alpha 把 Kolibri 权重完全开源了,Apache 2.0 许可随便用,78B 参数 MoE 架构还特意喂了 21% 德语数据。
A European lab reaching benchmark parity with Chinese and American open models in a bilingual German-English model suggests small-model training capability has spread well beyond the usual hubs. The Apache 2.0 license makes it directly usable for on-premise deployments, which matters for German public-sector and enterprise buyers with data-sovereignty requirements. Teams comparing open models should weigh its German-language edge against its weaker SWE-Bench Verified result.
Kolibri positions compliance as a design feature rather than a bolt-on, making it a candidate for European deployments where GDPR and AI Act alignment block US models. The Apache 2.0 license allows local hosting and fine-tuning without licensing fees, a practical option for ASEAN enterprises evaluating sovereign-AI alternatives to Mistral. Teams needing German-English bilingual capability gain a purpose-built tokenizer.