July 16, 2026
Benchmark? More like bench-war
German AI consortium releases Soofi S, an open 30B model that tops benchmarks
Europe’s new AI star drops, and the comments instantly turn into a benchmark brawl
TLDR: Germany’s Soofi S is a new open AI model that claims top scores while running efficiently on very long documents, marking a big European push in the AI race. But commenters immediately split into hype and suspicion, arguing over cherry-picked tests, training methods, and whether “open” really means open.
Germany’s new open AI model, Soofi S, was supposed to be a victory lap: a homegrown chatbot brain trained entirely on Deutsche Telekom’s cloud, built to handle giant text inputs without slowing to a crawl, and bold enough to claim it beats other openly available rivals in German, English, and coding tests. But the internet did what it does best: turned the launch into a noisy, glorious comment-section cage match.
A big chunk of the crowd wasn’t even talking about the model first — they were hyped that Europe had finally shown up with something serious. One commenter basically declared “Germans have entered the chat,” which is exactly the kind of meme-energy this launch sparked. Another said the real story might be the infrastructure, not the software, with excitement that this is one more non-US, non-China player entering the race.
Then came the skepticism. Critics rolled in accusing Soofi S of being “overtrained” — in plain English, fed an enormous mountain of text compared with its size — while others questioned whether its benchmark wins are flattering because newer rivals weren’t included. The sharpest knife came from a commenter claiming it only tops tests because the tests were in the training data, linking to a post on X. Even the phrase “fully open” got side-eye, with people debating whether “license-free for industry” is genuine openness or a clever legal magic trick. So yes, Germany launched an AI model — and the community launched a drama festival right after.
Key Points
- •A German consortium coordinated by KI Bundesverband released the open model Soofi S 30B-A3B, trained entirely on Deutsche Telekom’s Industrial AI Cloud in Munich.
- •Soofi S uses a hybrid Mixture-of-Experts architecture with 31.6 billion total parameters but activates only about 3.2 billion per token.
- •The model is reported to outperform fully open rivals such as OLMo 3 32B and Apertus 70B on German, English, and programming benchmarks.
- •The architecture reduces long-context bottlenecks by limiting KV-cache use to 6 of 52 layers, enabling nearly flat throughput from 4,000 to 256,000 tokens.
- •An update added criticism that the model appears overtrained under Chinchilla scaling laws, while project lead Michael Fromm argued those dense-model rules do not directly apply to MoE systems.