German AI consortium releases Soofi S, an open 30B model that tops benchmarks in both English and German
A German research consortium launched Soofi S 30B-A3B, an open 30-billion-parameter language model trained on Deutsche Telekom’s cloud infrastructure. It outperforms all fully open competitors on both English and German benchmarks.
- Soofi S 30B-A3B is an open 31.6B-parameter model trained on Deutsche Telekom’s cloud infrastructure in Munich.
- It outperforms all fully open competitors on both English and German benchmarks due to a German-weighted training dataset.
- The model uses a hybrid architecture that activates only a fraction of parameters per token, improving efficiency for long contexts.
- This release marks a notable advancement for European AI, emphasizing open-source accessibility and regional language optimization.
A German AI research consortium has unveiled Soofi S 30B-A3B, an open-source language model with 31.6 billion parameters, trained entirely on Deutsche Telekom’s cloud infrastructure in Munich. The model introduces a hybrid architecture designed to activate only a fraction of its parameters per token, enabling stable throughput even with very long contexts. Unlike many multilingual models that prioritize English, Soofi S was deliberately trained with a German-weighted dataset, allowing it to surpass all fully open competitors on both German and English benchmark tests.
The consortium behind Soofi S includes academic and industry partners, positioning the model as a significant step for European AI development. Its open release aligns with growing demand for high-performance, regionally optimized language models that balance efficiency with multilingual capability. The hybrid architecture also suggests potential applications in enterprise settings where long-context processing is critical, such as document analysis or conversational AI systems.
Provides a high-performance, open multilingual model optimized for German and English, with efficient long-context processing.
Offers a European-based alternative for AI applications requiring strong multilingual support and cloud-native deployment.
Highlights Europe’s growing role in open AI development with regionally tailored models.
- hybrid architecture
- A model design that activates only a subset of parameters per input token, reducing computational load while maintaining performance.
Open Source🇮🇳 ForThePeople.in: One Engineer’s Quest to Put Every Indian District’s Data in Your Pocket
Chinese AI startup MiniMax plans to open-source a 2.7 trillion parameter model later this year
Open SourceNVIDIA and Hugging Face Bring New Models and Frameworks to LeRobot for the Open Robotics Community
Open SourceTencent Releases Hy3: An Open 295B Mixture-of-Experts (MoE) Model with 21B Active Parameters and 256K Context
Open SourceLeRobot v0.6.0: Imagine, Evaluate, Improve
Civilian Protection in the Age of Military AI - Just Security
A new analysis argues military AI systems could inadvertently endanger civilians without stricter oversight and ethical safeguards.
AI ResearchStart Building for Agents, Not Just Humans
A developer argues that software should be designed with AI agents in mind, rather than just humans.
Analysis of 380 trillion AI tokens reveals how the technology is transforming financial markets - YaleNews
Yale researchers analyzed 380 trillion AI tokens to show how the technology is transforming financial markets.
GVA TECH survey shows 90%+ legal AI use, Copilot leads - International Business Times
A GVA TECH survey reveals over 90% of legal firms now use AI tools, with Microsoft Copilot leading adoption.
Meta AI Splurge Continues and J.P. Morgan Is Worried for the Stock - Barron's
Meta’s aggressive AI investment strategy has drawn scrutiny from J.P. Morgan, which warns the spending could pressure the company’s stock performance.
Military AR headsets get new AI-ready eye‑tracking windows from Syntec Optics - Stock Titan
Syntec Optics has developed AI-ready eye-tracking lenses for military AR headsets, enhancing real-time data processing and user interaction.