40
Soofi S: Sovereign German-English Foundation Model (huggingface.co)
submitted 2 months ago by to c/buyeuropean@feddit.uk
6 comments fedilink hide all child comments
 

Soofi S 30B-A3B is a sovereign, open-source Mixture-of-Experts (MoE) hybrid Mambaโ€“Transformer foundation model for German and English. Its hybrid design activates only 3B of 30B parameters per token and keeps the inference cache near-constant as context grows, giving it a decisive throughput advantage over dense models for long-context, high-concurrency deployment.

Pretrained on roughly 27 trillion tokens with deliberately up-weighted German, Soofi S matches dense 14โ€“27B models on aggregate English and German benchmarks, achieves the best code aggregates in both languages among 17 open base models, and outperforms every European sovereign baseline in our comparison โ€” including ones far larger in active parameters. Among fully open models, it obtains the highest English and German evaluation scores, ahead of Olmo 3 32B and Apertus 70B. Soofi S was built end-to-end on the German Industrial AI Cloud, a sovereign HPC-scale AI infrastructure operated by Deutsche Telekom in Munich.

...

you are viewing a single comment's thread
view the rest of the comments
[โ€“] 4 points 2 months ago (1 child)

From the page:

Consortium coordinated by the KI Bundesverband. Funded by the German Federal Ministry for Economic Affairs and Energy (BMWE) in the context of IPCEI-CIS and 8ra.

So Germany. But it's also bold of you to assume that "privately owned" means that it won't serve the interests of certain nation state actors.

Keeping Silicon Valley tf away from European public infrastructure and data is a must, since they not only represent the US gov - but also anyone else willing to pay them, like Russian or even Chinese nation state actors - adding to their problematic history of being hostile towards consumers and predatory against markets.

Europe knows this and is mobilising according to European interests and not those of random oligarchs and autocrats that want to define what AI is supposed to be.

All this being said, the weights aren't open - which is ofc something that will be a constant, since it would probably reveal a vast treasure trove of copyright infringement.

It's all mass surveillance and copyright infringement on stilts anyways, so yeah.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [โ€“] 4 points 2 months ago

    All this being said, the weights aren't open - which is ofc something that will be a constant, since it would probably reveal a vast treasure trove of copyright infringement.

    The source datasets seems closed (or I could not find them, perhaps deeper in the git repository). The weights are accessible through a custom license today and they intend to publish those after the beta.

    Apertus is the only larger model I know that published the datasets used in training.

  • permalink
  • fedilink
  • source
  • parent