AI Model Timeline

Tracking the accelerating release frequency of frontier AI models.

Soofi S 30B-A3B

Soofi Project·

LLMsopen-weightEdge / Consumer

Overview

Sovereign open-weight foundation model from the German SOOFI consortium (Fraunhofer IAIS/IIS, DFKI, TU Darmstadt, hessian.AI and others), coordinated by the KI Bundesverband and funded by the German Federal Ministry for Economic Affairs and Energy. Hybrid Mamba-Transformer MoE (~31.6B total, ~3.2B active), pretrained on ~27T tokens with up-weighted German and trained end-to-end on Deutsche Telekom infrastructure in Munich. Among fully-open models it reports the highest German and English aggregate scores (ahead of Apertus 70B, OLMo 3 32B, EuroLLM and Alia 40B), while still trailing broader open-weight leaders such as Qwen 3.5. Full transparency on data, training and evaluation. Vendor-reported code scores: HumanEval 73.8, MBPP 70.2, German-MBPP 84.2.

Capabilities and innovations

German + English focus (sovereign EU model)Runs locally (~3.2B active parameters)Full training/data/eval transparencyTrained end-to-end on EU infrastructureHybrid Mamba-2 / Transformer MoE (128+1 experts, 6 active per token)Strongest fully-open model on combined German+English benchmarksSovereign AI: trained entirely in Germany on Deutsche Telekom compute

Benchmarks

No benchmark values are recorded for this model.

Architecture and hardware

Parameters
31.6B total, 3.2B active per token (MoE)
Estimated VRAM at Q4
~19 GB, Edge / Consumer class
Quantization formats
GGUF
Recommended runtime
llama.cpp
License
Open (final license pending)

Links

Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.