European researchers release Soofi S 30B-A3B Mamba MoE trained on 27T tokens

New competitive European model (30bA3b) from Soofi with SYNTH inside. https://twitter.com/effi288/status/2075904321707798699

germans released a model that's actually not terrible it's small, and still worse than Qwen3.5, but it's very comparable to Nemotron 3 Nano but I guess that's what 27T pre-training tokens will do to a model https://twitter.com/effi288/status/2075904321707798699

For anyone wondering why they did not post-train this is actual sovereignty in the good sense of the world: slowly building strategic autonomy up to the pretraining level, while having the humility to start from the best practices in the open.… Show more

> 🇩🇪 Trained end-to-end on Deutsche Telekom’s Industrial AI Cloud in Munich — sovereign AI infrastructure on German soil Unexpected German W! https://twitter.com/effi288/status/2075904321707798699

new "sovereign" base model by a german consortium with a fun approach they take the exact same architecture (and most hyperparameters) as nemotron 3 nano and do a slightly different mixture of datasets, trained for ~26T tokens. the overlap with nemotron's mixture is roughly 80%… Show more