SHELFCRITTER AI Digest

MODEL

Soofi S

modeltopic-noteopen-sourceeurope

Overview

Soofi S is a fully-open 30B Mamba-Transformer MoE released by a German AI consortium on July 25, 2026 — 31.6B-total / 3.2B-active parameters, trained on 27T tokens across up to 512 B200 GPUs at Deutsche Telekom’s Industrial AI Cloud Munich (~253k GPU-hours). Positioned as a genuinely open (OSAID 1.0-compliant with ~99% of training data reconstructible) EU-sovereign alternative — compute, architecture, data, and weights are all EU-native. Beats OLMo 3 32B and Apertus 70B on English aggregates (70.1) and German (79.1), scores 73.8% on HumanEval. Second-tier vs frontier flagships on absolute benchmarks; the wedge is stack sovereignty, not benchmark leadership.

Timeline

  • 2026-07-25-AI-Digest — German consortium ships Soofi S as a fully-open 30B model — 31.6B-total / 3.2B-active Mamba-Transformer MoE, trained on 27T tokens across up to 512 B200 GPUs at Deutsche Telekom’s Industrial AI Cloud Munich (~253k GPU-hours). Beats OLMo 3 32B and Apertus 70B on English aggregates (70.1) and German (79.1); scores 73.8% on HumanEval. Meets Open Source AI Definition 1.0 — ~99% of training data is reconstructible, which is a stricter standard than “weights-only open” and closer to the Apertus / OLMo posture than the Mistral / Llama posture. Narrow read: second-tier vs Fable 5 and Opus 5 on absolute benchmarks — this is not a frontier model. What it is: a genuinely-open 30B MoE trained on sovereign-EU compute with reconstructible data, priced for local deployment. Structural read the corpus carries: the interesting axis here is not model quality but stack sovereignty — compute (Deutsche Telekom), architecture (Mamba-Transformer MoE), training data (reconstructible), and weights (open under OSAID 1.0) are all EU-native. That’s a distinct wedge from both the US frontier labs and the Chinese open-weight camp: not the cheapest, not the best, but the only stack that a European public-sector procurement can defend end-to-end without a US or Chinese dependency. Read as a procurement-ready alternative, not a benchmark-beater. First Soofi S entry in the corpus.

Key Developments

  1. First Fully-Open 30B EU-Sovereign Model (July 25, 2026): 31.6B-total / 3.2B-active Mamba-Transformer MoE trained on 27T tokens across up to 512 B200 GPUs at Deutsche Telekom’s Industrial AI Cloud Munich (~253k GPU-hours). OSAID 1.0 compliant with ~99% of training data reconstructible — stricter than “weights-only open” and matches the Apertus / OLMo posture. Beats OLMo 3 32B and Apertus 70B on English aggregates (70.1) and German (79.1); 73.8% HumanEval. The corpus framing: read as a procurement-ready alternative for European public sector, not a benchmark-beater against Fable 5 / Opus 5 — the wedge is stack sovereignty (compute + architecture + data + weights all EU-native), distinct from both US frontier labs and the Chinese open-weight camp.

See also: Apertus, Aleph Alpha, Hugging Face, MOC - Open Source Models.