German AI consortium releases Soofi S 30B

// 2h agoMODEL RELEASE

German AI consortium releases Soofi S 30B

A German research consortium released Soofi S 30B-A3B, an open language model trained entirely on Deutsche Telekom's cloud infrastructure. Featuring an efficient hybrid architecture and a dataset weighted toward German, it reportedly outperforms open competitors on German and English benchmarks.

// ANALYSIS

This release is a significant step forward for European AI sovereignty, demonstrating that competitive foundation models can be trained entirely on local infrastructure.

* It highlights the growing capability of European cloud providers, specifically Deutsche Telekom, to handle massive AI training workloads.

* The hybrid architecture emphasizes efficiency, addressing the common bottleneck of long-context processing throughput.

* Prioritizing bilingual (German/English) performance fills a crucial gap for the European enterprise market, which often struggles with models overly optimized for English.

// TAGS

aillmopen-sourcegermanysoofi-s

DISCOVERED

2h ago

2026-07-16

PUBLISHED

5h ago

2026-07-16

RELEVANCE

8/ 10

AUTHOR

amai

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

LAUNCH13m ago

Schema tops ARC-AGI-3 benchmark reasoning like physicists

Developed by Impossible Research, Schema is a custom agentic harness that structures LLM reasoning via inverse graphics and inverse dynamics. Guiding agents to reason like physicists, it achieved 99% Relative Human-Averaged Evaluation on the ARC-AGI-3 public set using Claude Opus 4.8 and Fable 5.

RESEARCH1h ago

Harness Handbook makes AI agent harnesses readable

The "Harness Handbook" is a newly released research paper (arXiv:2607.13285) that tackles the complexities of managing AI agent evaluation and deployment environments. It introduces approaches to improve the developer experience by ensuring that as agent harnesses evolve, they remain readable, easy to navigate, and straightforward to edit.

UPDATE1h ago

Pi v0.80.10 ships Kimi adaptive thinking, restores xAI

Pi v0.80.10 addresses several issues and introduces new capabilities, notably enabling Kimi Coding models to correctly use adaptive thinking, mirroring Anthropic's approach without token budgets. It also fixes a bug from v0.80.9 that removed xAI models from the catalog, corrects pricing metadata for Moonshot AI, and adds support for replaying empty-signature thinking blocks in K3.