Meet Soofi S 31.6B, the German 27T-Token Open MoE

Soofi S 31.6B is an open-weights foundation model with 31.6 billion total parameters and 3.2 billion active per token, developed by a German consortium and trained entirely on Deutsche Telekom's Industrial AI Cloud in Munich. It uses a hybrid Mamba-2 plus granular Mixture-of-Experts architecture, was pretrained on roughly 27 trillion tokens with a deliberate German-English focus, and supports a 1 million token context window. The model was released as a base checkpoint (no instruction tuning) alongside its full pretraining report on arXiv and weights on Hugging Face.

Key Takeaways

• Soofi S 31.6B has 31.6B total parameters with only 3.2B activated per token via a 128-expert MoE (6 active + 2 shared).
• The architecture stacks 52 layers: 23 Mamba-2 blocks, 23 MoE blocks, and 6 GQA attention blocks, adopting NVIDIA's Nemotron 3 Nano recipe unmodified.
• Pretraining ran on ~26.68–27T tokens across a three-phase Warmup-Stable-Decay schedule, with German upweighted to 15.32% during the anneal phase.
• Training consumed ~253,000 GPU-hours on up to 512 NVIDIA B200 GPUs at Deutsche Telekom's Munich site, powered by 100% renewable energy.
• Reported aggregate scores are 70.1 English and 79.1 German, leading fully open baselines and beating Olmo 3 32B and Apertus 70B on those suites.
• Weights, intermediate checkpoints, full data inventory, and eval code are gated-open on Hugging Face; the final commercial license is still being finalized.

What Is Soofi S 31.6B?

Soofi S 31.6B (officially Soofi S 30B-A3B) is a base language model built by the Soofi consortium, a group coordinated by the KI Bundesverband and including Fraunhofer IAIS/IIS, DFKI, TU Darmstadt, University of Würzburg, ellamind, and Merantix Momentum. The project was funded with roughly €20 million from Germany's IPCEI-CIS program and positioned as a sovereign European foundation for industrial and agentic use. It is a raw pretrained checkpoint without supervised fine-tuning or alignment, released alongside instruct and thinking previews codenamed Isar and Rhine.

The model surfaced publicly through an official pretraining report on arXiv (2607.09424, submitted July 10 and revised July 13, 2026) and a coordinated announcement by Dr. Michael Fromm, Head of Pretraining Data, on July 11. Coverage from Marktechpost and The Decoder followed on July 13 and 15, framing the release as the strongest fully open German-English open-weights model to date.

Soofi S 31.6B at a Glance

Field	Detail
Developer	Soofi consortium (KI Bundesverband-coordinated; Fraunhofer IAIS/IIS, DFKI, TU Darmstadt, Uni Würzburg, ellamind, Merantix Momentum)
Type	Base foundation model (hybrid Mamba-2 + granular MoE + GQA)
Total parameters	31.6B
Active parameters per token	3.2B
Architecture reference	NVIDIA Nemotron 3 Nano, adopted unmodified
Layers	52 total (23 Mamba-2, 23 MoE, 6 GQA)
Experts	128 experts, 6 active + 2 shared per token
Training tokens	~26.68–27T
Context window	1M tokens (Phase 3)
Knowledge cutoff	End of 2025
Training compute	Up to 512 NVIDIA B200 GPUs, ~253,000 GPU-hours
Training location	Deutsche Telekom Industrial AI Cloud, Munich (100% renewable)
Training window	March 24 – May 13, 2026
Modality	Text (German-English focus)
Weights	Gated preview on Hugging Face (Soofi-Project/Soofi-S-Base)
License	"Other" (custom, not yet finalized)
Official API	Not yet confirmed
Companion release	Isar (instruct), Rhine (thinking), GGUF + FP8 builds; Soofi L 120B-A12B in training

How Soofi S 31.6B Works

Soofi S 31.6B is what the consortium calls a Hybrid Mamba-MoE stack: 52 layers alternate 23 Mamba-2 state-space blocks with 23 granular MoE blocks, and only 6 layers use grouped-query attention. That ratio is why the model can hold a 1M-token window without paying the full quadratic cost of attention across every layer.

The MoE routing is what the report labels a Granular MoE Layer — 128 experts per MoE block, with 6 experts activated plus 2 shared per token. That gives a 3.2B active-parameter budget while the total sits at 31.6B, and the architecture is copied unmodified from NVIDIA's Nemotron 3 Nano recipe. Pretraining followed a three-phase WSD Schedule (Warmup-Stable-Decay): ~20T breadth tokens in phase one, ~6.58T high-quality anneal tokens with German pushed to 15.32% in phase two, and ~0.1T long-context tokens in phase three to lock in the 1M window.

Two efficiency claims stand out. The report says decode throughput per GPU runs 8–9× dense 14–24B models at a 40k-token context and batch size 32, and stays flat from 4k up to 256k. Second, the whole training run consumed roughly 253,000 GPU-hours on renewable-powered infrastructure — an unusually detailed sovereignty-and-sustainability disclosure for a model of this size.

What You Can Do With Soofi S 31.6B

The consortium is targeting Sovereign Industrial Agents: long-document analysis, code, and multi-step tool-using pipelines that European industrial partners can host on-premise. Because Soofi S ships as a base model, most current use cases involve downstream fine-tuning, and ellamind is publishing evaluation spaces to support that workflow.

Early community traction skews toward local inference. GGUF quantized builds are already circulating, with users reporting Mac Studio-class hardware running the model locally. The instruct and thinking previews (Isar and Rhine) give developers a starting point for chat and reasoning fine-tunes without waiting for a formal SFT release. If you're building similar bilingual open-weights workflows and want an alternative to test alongside, kie.ai catalogs models like Claude Sonnet 5 for direct chat comparison.

For teams tracking European AI sovereignty specifically, the full data inventory and hyperparameter disclosure make Soofi S 31.6B one of the few 30B-class models where you can audit exactly what went into pretraining — a decisive practical difference from most closed frontier releases.

How Soofi S 31.6B Compares

Model	Total params	Active params	English aggregate	German aggregate
Soofi S 31.6B	31.6B	3.2B	70.1	79.1
Olmo 3 32B (fully open)	32B	32B	67.3 (per report)	Not yet confirmed
Apertus 70B (fully open)	70B	70B	Not yet confirmed	72.8 (per report)
Qwen3.5 (open-weights)	Not yet confirmed	Not yet confirmed	Leads on some benches	Not yet confirmed

According to the official report, Soofi S beats Olmo 3 32B by 2.8 points on the English aggregate and Apertus 70B by 6.3 points on the German aggregate, both while activating a fraction of the parameters. Specific benchmark scores include 73.8 HumanEval, 84.2 MBPP-DE, 88.8 GLP-DE, 61.2 INCLUDE-DE, 86.1 GSM8K, 56.0 Minerva MATH-DE, and 43.4 GPQA-Diamond. Independent commentators, including a critical analysis by community reviewer @0xGhostnt, note the model trails top Chinese open-weights like Qwen3.5 on some benchmarks and cannot be positioned as frontier-tier.

For readers tracking parallel foundation-model launches, our earlier coverage of the Grok 4.5 foundation traces gives useful context on how closed-frontier training footprints compare with Soofi S's fully documented run.

Availability: How to Access Soofi S 31.6B

The primary access path is Hugging Face. The base checkpoint lives at Soofi-Project/Soofi-S-Base as a gated preview, alongside intermediate checkpoints, full training data inventory, hyperparameters, and evaluation code. Instruct and thinking previews (Isar and Rhine) and quantized GGUF and FP8 builds are also published under the same organization.

The consortium's project site is soofi.info. There is no official first-party API yet, and the license is marked "other" pending finalization — meaning production commercial use is on hold until the consortium publishes the final license text. Early download counts on Hugging Face were reported in the mid-double-digits within the first week.

What We Don't Know Yet

Several material items remain open:

• Final commercial license text. The current "other" designation and gated preview status mean production use terms are not yet settled.
• Independent long-context verification. The official report itself flags RULER extraction gaps, and no third-party 1M-context stress tests are public.
• Hosted API pricing. No first-party API or per-token pricing has been announced; the consortium's stated focus is research and fine-tuning.
• OSAID 1.0 status durability. The report says the release meets OSAID 1.0, but ~1.3% of training data comes from a commercial Genios exception — how that survives stricter EU open-data proposals is unclear.
• Frontier-tier comparisons. Public head-to-heads against closed frontier models (Claude, GPT, Gemini) or larger Chinese open-weights in real production workloads have not been published.

Frequently Asked Questions

What is Soofi S 31.6B?

Soofi S 31.6B is an open-weights foundation model with 31.6 billion total parameters and 3.2 billion active per token, built by a German consortium coordinated by the KI Bundesverband. It uses a hybrid Mamba-2 plus Mixture-of-Experts architecture and was pretrained on roughly 27 trillion tokens focused on German and English.

Is Soofi S 31.6B open source?

Soofi S 31.6B ships with open weights, intermediate checkpoints, full data inventory, hyperparameters, and training and evaluation code on Hugging Face. The final commercial license text is listed as "other" and, based on the official release notes, is not yet finalized for unrestricted production use.

Who made Soofi S 31.6B?

Soofi S 31.6B was built by the Soofi consortium, coordinated by the KI Bundesverband and including Fraunhofer IAIS/IIS, DFKI, TU Darmstadt, University of Würzburg, ellamind, and Merantix Momentum. It was funded with roughly €20 million from Germany's IPCEI-CIS program and trained on Deutsche Telekom's Industrial AI Cloud in Munich.

How was Soofi S 31.6B trained?

Soofi S 31.6B was trained end-to-end from March 24 to May 13, 2026 on up to 512 NVIDIA B200 GPUs, consuming roughly 253,000 GPU-hours on 100% renewable energy. Training used a three-phase Warmup-Stable-Decay schedule spanning ~20T breadth tokens, ~6.58T high-quality anneal tokens, and ~0.1T long-context tokens.

What is the context window of Soofi S 31.6B?

Soofi S 31.6B supports a 1 million token context window, reached during the third training phase dedicated to long inputs. Independent verification on stress tests like RULER extraction is still limited, and the official report flags gaps at the far end of the window.

How does Soofi S 31.6B compare to other open models?

Soofi S 31.6B scores 70.1 on an aggregate English benchmark suite and 79.1 on German, leading fully open baselines. According to the official report, it beats Olmo 3 32B by 2.8 points in English and Apertus 70B by 6.3 points in German, but trails top Chinese open-weights such as Qwen3.5 on some benchmarks.

Can I download Soofi S 31.6B?

Soofi S 31.6B weights are available on Hugging Face under the Soofi-Project organization as a gated preview, alongside intermediate checkpoints and quantized GGUF and FP8 builds. There is no official hosted API yet, and access is centered on the consortium's Hugging Face repositories and the soofi.info project site.

What to Watch Next

Three signals will shape whether Soofi S 31.6B graduates from research milestone to production tool: publication of the final commercial license text, independent long-context and agentic benchmarks from parties outside the consortium, and the release timeline for Soofi L (the 120B-A12B follow-on the consortium says is already training). Each will materially change the model's position against both fully open baselines and Chinese open-weights competitors.

Building similar bilingual chat and reasoning workflows? On kie.ai you can try Claude Sonnet 5, Gemini 3 Pro, and GPT-5.6.