• 4 min read
Soofi S claims top open scores with just 3.2B active params
Germany’s Soofi S is a 30B open model trained in Munich that leads open English and German benchmarks while keeping long-context speed unusually flat.
Image: Hacker News
A German research consortium says its new Soofi S 30B-A3B is now the top-performing fully open model on both English and German benchmark aggregates, beating earlier open leaders including OLMo 3 32B and Apertus 70B. The model was trained entirely on Deutsche Telekom’s Industrial AI Cloud in Munich and leans heavily on German-language data.
An update published on July 15, 2026 addressed criticism that Soofi S was heavily overtrained relative to classic Chinchilla scaling laws from Google DeepMind. Those 2022 laws suggest roughly 20 tokens per parameter for a fixed compute budget. Soofi S was trained on about 27 trillion tokens with 30 billion parameters, far above that ratio. Counting only the 3.2 billion parameters active per token pushes the ratio even higher. Michael Fromm, part of the project’s technical leadership, argued that those dense-model rules do not transfer cleanly to Mixture-of-Experts systems and pointed to Nvidia, which trained its own models on up to 25 trillion tokens.
Soofi S uses a hybrid Mamba-Transformer design based directly on Nvidia’s Nemotron 3 Nano architecture. It has 31.6 billion parameters in total, but activates only about 3.2 billion for each generated token, giving it a compute profile closer to a 3B model than a standard 30B system.
That architecture is the core of the pitch. Only 6 of Soofi S’s 52 layers maintain a KV cache, which reduces the usual long-context bottleneck. At 40,000 tokens of context with 32 parallel requests, the consortium says Soofi S produces roughly eight times more tokens per second per GPU than dense models in the 14 to 24 billion parameter range. Its throughput also stays nearly flat from 4,000 to 256,000 tokens. The only model showing similar behavior in the reported tests is Alibaba’s Qwen3.5 35B-A3B.
Training data and German weighting
The training run was split into three phases and covered about 27 trillion tokens overall:
Recommended reading
Hassabis says STEM makes you 10x better at AI
- roughly 20 trillion tokens of broad web, code, math, and domain data
- about 6 trillion tokens of higher-quality sources
- a final phase on very long documents up to one million tokens
German was a deliberate focus. It made up 7.2 percent of the mix in phase one and 15.3 percent in phase two. By comparison, the Nemotron reference recipe assigns only about 5 percent to all non-English languages combined.
The German data came from sources including HPLT, the openly licensed German Commons corpus, German portions of FinePDFs and FineWiki, plus the commercially licensed Genios corpus with 193 million newspaper articles from 916 German publications. The mix also included machine-translated and synthetic German texts.
Benchmarks, weaknesses, and release terms
In the reported evaluations against 16 other open models, Soofi S leads all fully open models on aggregate scores in both languages.
The model scored 70.1 on the English aggregate and 79.1 on the German aggregate. On code tasks, it reached 73.8 percent on HumanEval, 70.2 on MBPP, and 84.2 on the German MBPP variant. On INCLUDE-DE, it tied Qwen3.5 35B-A3B at 61.2 points.
Compared with the Nemotron baseline, the consortium says the German-focused data recipe improved language proficiency by 15.1 points and GPQA-Diamond by 9.6 points, without hurting English performance.
The model is not uniformly stronger. It scored 56 on Minerva MATH-DE, behind Qwen3.5 35B-A3B at 76.5 and Gemma 3 27B at 65.6, and it trails on NaturalQuestions open factual retrieval.
The RULER long-context test exposed a sharper weakness. On a task requiring extraction of frequently occurring words from long text, Soofi S falls to around 3 percent hit rate beyond 32,000 tokens, while the comparable Nemotron model still reaches 60 to 64 percent. The authors blame the lack of synthetic extraction-focused data in long-context training.
Training ran from March to May on up to 512 Nvidia B200 GPUs, using about 253,000 GPU-hours. The project says the Munich facility runs on renewable energy, uses Eisbach canal water for cooling, and feeds waste heat into the Tucherpark neighborhood.
The consortium is coordinated by the German AI Association and funded by the German Federal Ministry for Economic Affairs and Energy under the European IPCEI-CIS program. Participants include Fraunhofer IAIS, Fraunhofer IIS, DFKI, TU Darmstadt, the University of Würzburg, the L3S Research Center, the Berlin University of Applied Sciences, Ellamind, and Merantix Momentum.
The team is releasing the model weights, selected intermediate checkpoints, full training and evaluation code, and a detailed data inventory. It says Soofi S meets the Open Source AI Definition 1.0 from the Open Source Initiative, though it does not satisfy a stricter proposed European open-data standard because 1.3 percent of the training mix comes from Genios under commercial license. The exact release license has not yet been finalized.
AI Editor
Ava covers the rapidly evolving world of artificial intelligence, from foundational models and research labs to the real-world economics of intelligence. With a background in computational linguistics, she cuts through the hype to find out what actually works. She firmly believes that benchmarks are just marketing until reproduced in the wild.
via Hacker News
