UPDATE: The Soofi S Tech Report v3.0 is now online (arxiv.org/pdf/2607.09424). New analyses, a revised evaluation, and 152,000 individual results that anyone can verify. Architecture & Data Soofi S uses the openly published reference architecture of Nemotron 3 Nano (cf. Tech Report, p.1), which allowed us to train a capable first model within a few months. This is how the field works today: Kimi K2 builds on DeepSeek, many models (including from Mistral) build on Llama. Architectures are published knowledge that the open-source community should be using collectively. In parallel, we are researching our own architectures as a further building block of a sovereign European model stack. Where does the difference lie? Soofi S makes the difference with data: our data recipe delivers substantial gains in German in particular (+15 points in language proficiency vs. the architecturally identical reference model, see linked post), whilst also improving English performance. Same blueprints, better house. Practice exams and what the community found in our data Like other open projects (e.g. OLMo), we train with rephrased practice questions from well-known benchmarks: mock exams, not the real test, so that the model can practise the later test formats. The practice questions come from the training splits of the benchmarks, which are usually separated into "practice" and "exam". For GPQA and GPQA Diamond, however, this separation does not exist: the complete exam set is simply labelled "Train" on Hugging Face. As a result, exam questions ended up in the practice set, and from there in our data and presumably in other open-source models as well. The advantages of an open approach The community spotted this in our published data, and we are grateful for that. We verified the issue, removed GPQA and GPQA Diamond from Soofi S's scoring, and recalculated all comparisons across all 16 models. For verification: all 152,000 individual results for Soofi S are available here (huggingface.co/datasets/Soof…), for the other reference models they are available here (huggingface.co/datasets/ella…). Additionally, the linked post includes an evaluation by ellamind on unpublished held-out data that confirms the results (huggingface.co/datasets/ella…). Putting the data volume in perspective If the entire training corpus of Soofi S were a path around the Earth, the data in question would account for 68 centimetres. Issues like these can only be found when everything is in the open. The outcome shows: open source works. What comes next The post-trained models are currently in beta testing with partners from academia and industry. After that, the hardened model is to be released under a permissive licence. Interest has been considerable, and we are working at full pace towards an open release in the coming weeks. Thank you for your patience :) Many thanks also to everyone who looks this closely. This kind of scrutiny is precisely the point of open models. And there is more to come: Soofi L is already in training. Built by a consortium of @DFKI,@FraunhoferIIS @FraunhoferIAIS ,@TUDarmstadt,@Uni_WUE,@BHT_bln, L3S,@LamarrInstitute,@ellamindAI,@Hessian_AI and @MMerantix— coordinated by the@ki_verbandand funded by the BMWE (IPCEI-CIS / 8ra).