German Open-Source Foundation Model Soofi S Released
eliebakouch · x · 2026-07-12
The open foundation model project for German and English, Soofi S, released a pre-training technical report, emphasizing "sovereign model" and full transparency: full training report, data source breakdown, hyperparameters, training and evaluation code, checkpoints, and strong performance on both English and German aggregated benchmarks. The model was trained on Deutsche Telekom's infrastructure, with deliberate weighting on German data.
An additional comment notes that the experiment uses nearly the same architecture and most hyperparameters as Nemotron 3 Nano, with a slightly different data mix, trained on 26T tokens, with 80% overlap with Nemotron recipe. The commenter says it's not computationally efficient—continuing pre-training from Nemotron base would save compute—but it's a valuable case study: observing how different data mixes affect downstream evals, and demonstrating the practical value of pre-training from scratch for a European team.
Related event: German Alliance Releases Open-Source Sovereign Model Soofi S 30B-A3B(5 posts)→
More from Models
- Artificial Analysis breaks down all sub-evaluations in Intelligence Index v4.3.2 — ArtificialAnlys · 2026-10-27
- GPT-6.1 Terra gains ~20 points on Terminal-Bench 4.0, rivalling GPT-6.1 Sol — ArtificialAnlys · 2026-10-27
- GPT-6.1 Terra cuts hallucination rate by 50 points while lifting accuracy 2 points — ArtificialAnlys · 2026-10-27
- GPT-6.1 Terra scores 48 on AA Coding Agent Index at one-third the cost of GPT-5.6 Terra — ArtificialAnlys · 2026-10-27
- Leaked Terminal-Bench 4.0 chart shows GPT-6.1 Terra leading at lowest cost, unverified — Angaisb_ · 2026-10-27
- OpenAI brings GPT-6.1 Terra to the API and Codex for fast, cheap computer use — OpenAIDevs · 2026-10-27