Sovereign Open-Source DE/EN Model Soofi
The Soofi-Team · hf · 2026-07-13
This post introduces Soofi S 30B-A3B: a sovereign open-source foundation model tailored for German and English, built upon a MoE + Hybrid Mamba Transformer architecture.
While the model features 30B total parameters, it only activates 3B parameters per token. This ensures that the inference cache remains nearly constant in long-context and high-concurrency scenarios, delivering superior throughput compared to dense models. The authors note it was pre-trained on roughly 27 trillion tokens with a deliberately increased weight for German data. On comprehensive English-German benchmarks, it matches the performance of 14B to 27B dense models, and it ranks among the top 17 open foundation models in coding evaluations across both languages, outperforming all compared European sovereign baselines.
The project also strongly emphasizes a "fully open-source" delivery: model weights, selected intermediate checkpoints, complete data source statistics, hyperparameters, alongside training and evaluation code, will all be made publicly available. Training was conducted on the German Industrial AI Cloud.
Related event: German Alliance Releases Open-Source Sovereign Model Soofi S 30B-A3B(5 posts)→
More from Infra
- Lightning AI’s LitLogger captures training metrics, artifacts, commands, and environment data — LightningAI · 2026-07-21
- Moonshot pauses Kimi K3 signups five days after launch as GPU demand surges — eyishazyer · 2026-07-21
- Microsoft expands Mistral models across Azure, Foundry, Copilot Studio and Azure Local — arthurmensch · 2026-07-21
- SmolVM is pitched as a lighter in-house sandbox for agent runtimes — aniketmaurya · 2026-07-21
- Three-part PyTorch profiling series explains torch.profiler for accelerator debugging — RisingSayak · 2026-07-21
- AMD shows Ryzen AI Halo as a 100% local AI platform for on-device workflows — Sam Witteveen · 2026-07-21