German Research Consortium Unveils Open-Source AI Model Soofi
The German Association for Artificial Intelligence has officially launched Soofi S30B-A3B, a new open-source large language model. This milestone advances Europe’s sovereign AI infrastructure and revitalizes the open-source sector with a focus on high performance and efficiency.
Soofi S30B-A3B features an innovative architecture that balances efficiency and power. Utilizing a Mixture of Experts (MoE) design combined with Mamba-2 and standard attention layers, the model contains 31.6 billion parameters but activates only ~3.2 billion per token. This approach drastically cuts computational costs, delivering speeds comparable to smaller models while maintaining high throughput. Notably, it handles long contexts (up to 40,000 tokens) with eight times the generation throughput of traditional dense models, ensuring stability even as context length grows.

Training prioritized German language proficiency. Over 27 trillion tokens, the team increased German data proportion from 7.2% to 15.3%, integrating diverse sources like Wikipedia, news, and technical documents. This strategy ensures exceptional performance in complex German-language tasks.
Benchmarks confirm Soofi S30B-A3B’s leadership among open-source models in English and German, outperforming OLMo332B and Apertus70B. It excels in code generation, professional reasoning, and regional knowledge. However, the team notes room for improvement in ultra-long context extraction and German mathematical reasoning.

Trained on Deutsche Telekom’s Industrial AI Cloud in Munich using 512 Nvidia B200 GPUs, the project is part of the European IPCEI-CIS initiative. Powered entirely by renewable energy, the model’s weights, evaluation code, and data lists will be gradually open-sourced. This transparency supports European industrial applications, and the team is currently seeking industry partners to explore uses in technical documentation and code generation.
Related article
Toyota’s $6.4bn robotics estimate puts physical AI in focus
Toyota Motor projects that rolling out automation across its manufacturing facilities, subsidiaries, and key suppliers may require approximately 400,000 robots and an annual expenditure of roughly 1 trillion yen ($6.4 billion) starting in 2028.Accord
Meta Open-Sources 30B-Parameter Muse Glimmer for Local AI Agent on Consumer Hardware
On August 10, Meta launched Muse Glimmer, an open-source model designed to empower local AI agents on consumer hardware, marking a key step in its "Personal Superintelligence" strategy. With 30 billion parameters, this model is the open-source counte
GPT-5.6 May Launch Next Week, Focusing on Agent-Level Operational Capabilities
The competitive landscape of artificial intelligence is intensifying rapidly. Recent leaks regarding test updates have brought the new flagship model, GPT-5.6, into the spotlight. Market consensus suggests an official launch in the latter half of Jun
Related Special Topic Recommendations
Comments (0)
0/500
The German Association for Artificial Intelligence has officially launched Soofi S30B-A3B, a new open-source large language model. This milestone advances Europe’s sovereign AI infrastructure and revitalizes the open-source sector with a focus on high performance and efficiency.
Soofi S30B-A3B features an innovative architecture that balances efficiency and power. Utilizing a Mixture of Experts (MoE) design combined with Mamba-2 and standard attention layers, the model contains 31.6 billion parameters but activates only ~3.2 billion per token. This approach drastically cuts computational costs, delivering speeds comparable to smaller models while maintaining high throughput. Notably, it handles long contexts (up to 40,000 tokens) with eight times the generation throughput of traditional dense models, ensuring stability even as context length grows.

Training prioritized German language proficiency. Over 27 trillion tokens, the team increased German data proportion from 7.2% to 15.3%, integrating diverse sources like Wikipedia, news, and technical documents. This strategy ensures exceptional performance in complex German-language tasks.
Benchmarks confirm Soofi S30B-A3B’s leadership among open-source models in English and German, outperforming OLMo332B and Apertus70B. It excels in code generation, professional reasoning, and regional knowledge. However, the team notes room for improvement in ultra-long context extraction and German mathematical reasoning.

Trained on Deutsche Telekom’s Industrial AI Cloud in Munich using 512 Nvidia B200 GPUs, the project is part of the European IPCEI-CIS initiative. Powered entirely by renewable energy, the model’s weights, evaluation code, and data lists will be gradually open-sourced. This transparency supports European industrial applications, and the team is currently seeking industry partners to explore uses in technical documentation and code generation.
Toyota’s $6.4bn robotics estimate puts physical AI in focus
Toyota Motor projects that rolling out automation across its manufacturing facilities, subsidiaries, and key suppliers may require approximately 400,000 robots and an annual expenditure of roughly 1 trillion yen ($6.4 billion) starting in 2028.Accord
Meta Open-Sources 30B-Parameter Muse Glimmer for Local AI Agent on Consumer Hardware
On August 10, Meta launched Muse Glimmer, an open-source model designed to empower local AI agents on consumer hardware, marking a key step in its "Personal Superintelligence" strategy. With 30 billion parameters, this model is the open-source counte
GPT-5.6 May Launch Next Week, Focusing on Agent-Level Operational Capabilities
The competitive landscape of artificial intelligence is intensifying rapidly. Recent leaks regarding test updates have brought the new flagship model, GPT-5.6, into the spotlight. Market consensus suggests an official launch in the latter half of Jun





Home






