🤖Soofi S Debuts: Open-Source German Language Model
German AI Just Got a Major Boost
TL;DR
A German research group releases Soofi S, an open-source language model that outperforms rivals in both English and German benchmarks. It's trained to handle long inputs efficiently, making it a game-changer for multilingual tasks.
Soofi S is the latest addition to the AI landscape, boasting impressive performance across multiple languages. Trained on Deutsche Telekom’s cloud infrastructure with a focus on efficiency, this model stands out due to its hybrid architecture that activates only 3.2 billion parameters per token from a total of 31.6 billion. This makes it ideal for handling very long inputs without sacrificing speed or quality. Developers working in multilingual environments will find Soofi S particularly useful. Its superior performance on German benchmarks and English tasks, along with its ability to maintain throughput at high-context lengths, positions it as a strong contender for applications requiring robust language processing capabilities. The model was trained using 27 trillion tokens across three phases, with the data mix shifting towards higher-quality sources and more German-language content. It achieved scores of 73.8% on HumanEval and 84.2% on the German MBPP variant, significantly outperforming previous leaders like OLMo 3 32B and Apertus 70B.
Key Points
Soofi S, trained on Deutsche Telekom’s AI cloud, uses a hybrid architecture activating only 3.2 billion parameters per token from its total of 31.6 billion.
Training involved 27 trillion tokens across three phases, shifting towards higher-quality and more German-language data.
The model scores 73.8% on HumanEval and 84.2% on the German MBPP variant, outperforming OLMo 3 32B and Apertus 70B in benchmarks.
Soofi S generates roughly eight times more tokens per second per GPU than dense models in the 14 to 24 billion parameter range.
Training took place between March and May on up to 512 Nvidia B200 GPUs at Deutsche Telekom’s Industrial AI Cloud, totaling about 253,000 GPU-hours.
Why It Matters
If you're working with multilingual datasets or applications requiring robust German language processing, Soofi S is a game-changer. Its superior performance on both English and German benchmarks makes it ideal for tasks that demand high accuracy and efficiency. For teams focused on natural language understanding in German-speaking regions, this model offers significant advantages over existing solutions.
Frequently Asked Questions
Why does this matter?
If you're working with multilingual datasets or applications requiring robust German language processing, Soofi S is a game-changer. Its superior performance on both English and German benchmarks makes it ideal for tasks that demand high accuracy and efficiency. For teams focused on natural language understanding in German-speaking regions, this model offers significant advantages over existing solutions.
What happened?
A German research group releases Soofi S, an open-source language model that outperforms rivals in both English and German benchmarks. It's trained to handle long inputs efficiently, making it a game-changer for multilingual tasks.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 2,077 builders reading daily.