Soofi S 30B: German-focused Nemotron MoE for high-throughput long-context tasks, but weak extraction

A German open model that prioritizes German text and long documents, with important limits

Soofi S 30B‑A3B is a new open‑weight language model from a German research consortium coordinated by the KI Bundesverband. Trained on Deutsche Telekom’s Industrial AI Cloud in Munich, the model pairs a Nemotron‑style hybrid mixture‑of‑experts (MoE) architecture with a deliberate German‑heavy training mix. The payoff: unusually high tokens/sec at extreme context lengths and strong open‑model benchmark results for code and German language tasks, but also clear gaps in math reasoning and targeted extraction from very long documents.

What Soofi S is, in plain terms

According to the consortium’s pretraining report (arXiv:2607.09424v1), Soofi S is a 31.6 billion‑parameter model that uses a hybrid MoE design inspired by Nvidia’s Nemotron 3 Nano. The architecture deliberately activates far fewer parameters per token, the team reports about 3.2 billion active parameters per token, meaning roughly 3.2B parameters the router engages for each generated token rather than running the full 31.6B every time.

Pretraining totaled about 27 trillion tokens across three phases: Phase 1 ≈ 20T (broad mix), Phase 2 ≈ 6T (higher‑quality sources, heavier German share), and a Phase 3 that focused on very long documents. The report states Phase 3 exposed the model to documents up to 1, 000, 000 tokens; readers should consult the paper’s appendix for the length‑bucketing details and exact sampling scheme. The pretraining report is dated July 13, 2026 and notes the main training window occurred “between March and May” (see the report for confirmation of the calendar year and the infrastructure appendix for exact timing).

Key technical facts (as reported)

Model: Soofi S 30B‑A3B, 31.6 billion parameters.
Active parameters per token: ≈ 3.2 billion (i.e., the parameters the MoE router engages per token, reported in the architecture section).
Pretraining volume: ≈ 27 trillion tokens (Phase 1 ≈ 20T; Phase 2 ≈ 6T; Phase 3 long‑document stage).
German data share: Phase 1 = 7.2%; Phase 2 = 15.3% (the team intentionally up‑weighted German in later phases).
Commercial data fraction: Genios (a commercial German news archive) comprises 1.3% of the training mix.
Compute & infra (reported): training used up to 512 Nvidia B200 GPUs and about 253, 000 GPU‑hours (see the infrastructure appendix and model card for hardware/time breakdowns).
Release artifacts: model weights, selected checkpoints, training/eval code and a detailed data inventory are published (Hugging Face collection referenced in the report).

Performance highlights, and the caveats that matter

The consortium’s evaluation places Soofi S near the top of the open‑weight field on several English and German benchmarks. Representative figures from the pretraining report (evaluation conducted with the project’s harness and prompt settings) include:

HumanEval: 73.8% (per the report’s evaluation harness and prompt configuration).
MBPP: 70.2%.
German MBPP variant: 84.2%.
INCLUDE‑DE (Germany‑regional knowledge): 61.2 points (tie with Alibaba Qwen3.5 35B‑A3B in the report’s comparison set).
Minerva MATH‑DE (German competition math): 56, notably below Qwen3.5 (76.5) and Gemma 3 (65.6) in the same tables.
RULER long‑context extraction: Soofi S’s hit rate drops to ≈3% beyond roughly 32, 000 tokens; the report contrasts this with a Nemotron reference that scores ≈60-64% on the same task and attributes Soofi S’s failure to a lack of extraction‑targeted synthetic pretraining data.

Two patterns stand out. First, Soofi S is strong on code‑generation proxies and German‑language code variants, which matches the Phase‑2 German emphasis. Second, long‑context throughput does not equal precise extraction. The model can process very long inputs quickly but struggles to pull out specific facts unless it has been trained on extraction tasks.

Why the Nemotron‑style MoE matters for businesses

MoE designs keep many experts but route each token through a subset of them. That lowers FLOPs per token and can dramatically increase tokens/sec at long contexts. The report’s throughput tests show Soofi S producing roughly 8× more tokens/sec per GPU than dense 14-24B models under a 40, 000‑token context with 32 parallel requests (see the throughput figures and experimental conditions in the paper). The authors also report near‑flat generation throughput from 4, 000 tokens up to 256, 000 tokens in their setup, which is useful for agents and document workflows that need sustained throughput over long inputs.

Important operational caveats:

The paper describes Soofi S as “Nemotron‑style” with deliberate departures in data mix and curriculum, it is not an unmodified drop‑in of Nvidia’s Nemotron 3 Nano.
Some implementation details that matter for deployment (for example, exact KV‑cache layer counts and per‑layer caching strategy) are best verified in the released model code and model card; the publicly posted artifacts are the place to confirm those specifics.
High throughput is hardware‑and harness‑sensitive. Replicate the report’s measurement settings (GPU type, batch/parallelism, driver/runtime) when validating tokens/sec for your stack.

Openness, licensing and reproducibility

The consortium published weights, selected checkpoints, training and evaluation code, plus a detailed data inventory and states compliance with the Open Source AI Definition 1.0. They also report that about 99% of the training mix can be reconstructed independently, with 1.3% of the mix (the Genios corpus) under commercial license. Because of that commercial fraction, the release does not meet a stricter EU open‑data interpretation unless licensing changes; the final model license had not been finalized in the pretraining report, check the Hugging Face collection and model card for the license that ships with the weights.

Also note: the report says training occurred on Deutsche Telekom’s Industrial AI Cloud in Munich; operational claims about renewable energy, specific cooling sources or waste‑heat reuse appear in project communications but should be confirmed with Deutsche Telekom or the published model card before using them for procurement or sustainability claims.

Where Soofi S makes sense for business use today

For European enterprises and teams building long‑document agents, Soofi S offers three practical advantages:

German fidelity: A Phase‑2 German share of 15.3% and strong German MBPP performance make it a solid base model for localized documentation, support, and code‑generation tasks.
Sovereign infrastructure alignment: Training and release tied to Munich‑based infrastructure aligns with EU data‑sovereignty preferences and on‑prem deployment strategies.
Long‑context cost/latency profile: High tokens/sec at extreme contexts can reduce GPU costs and latency for RAG, multi‑document summarization, or agent workflows that read long manuals or logs.

Don’t assume throughput equals out‑of‑the‑box correctness. The RULER results show that accuracy on extraction tasks degraded dramatically beyond ~32k tokens. For production deployments you should treat Soofi S as a promising foundation that will often need system‑level mitigations:

retrieval and index layers to reduce the need for scanning entire documents;
targeted fine‑tuning or synthetic extraction training to recover pointer/lookup behavior; and
runtime result validation for high‑stakes outputs (legal, financial, code execution).

Three‑step POC plan for technical teams

License and provenance check: Confirm the final license on the Hugging Face collection and inspect the published data inventory (verify the 1.3% Genios fraction and the report’s ~99% reconstructibility statement).
Targeted capability tests: Run throughput and extraction benchmarks with your hardware and serving stack. Include RULER‑style extraction tests at the context lengths you expect in production, plus end‑to‑end RAG latency and downstream QA metrics.
Safety and verification: Perform hallucination, bias and red‑team checks on your domain prompts; for code generation, add execution and unit tests before any automated deployment.

Practical checklist before deployment

Confirm the model license and commercial‑use terms in the published model card and Hugging Face collection.
Replicate the consortium’s throughput tests on your hardware and measure tokens/sec, latency and cost per useful token.
Run RULER or custom extraction datasets at your target context lengths; don’t assume long‑context training implies accurate extraction.
Request or run safety/red‑team audits; the pretraining report documents technical evaluations but does not include full deployment‑grade safety audits.
If sustainability or infrastructure provenance matters, verify operational claims about the Munich facility with Deutsche Telekom or consortium PR materials.

Key questions, short, honest answers

Is Soofi S truly “open”?

The consortium published weights, checkpoints, training/eval code and a data inventory and claims compliance with the Open Source AI Definition 1.0. However, 1.3% of the training mix (the Genios news archive) is commercially licensed, so a stricter EU open‑data interpretation is not met; confirm the final license on the Hugging Face collection before relying on the model for commercial products.
Will Soofi S save inference cost for long documents?

In the consortium’s experiments, yes: the report shows roughly 8× more tokens/sec per GPU than dense 14-24B models at a 40, 000‑token context with 32 parallel requests and near‑flat throughput up to 256k tokens. Reproduce the test with your hardware and serving stack to validate savings for your workload.
Is it reliable for high‑end math or competition problems?

No, not relative to some peers. On Minerva MATH‑DE Soofi S scores 56 in the report, below Qwen3.5’s 76.5 and Gemma 3’s 65.6. Expect to need targeted fine‑tuning or a different model for advanced math reasoning.
Does long‑context training mean accurate extraction from huge documents?

Not automatically. The RULER benchmark shows Soofi S’s extraction hit rate falls to about 3% beyond ~32k tokens; the team attributes this to limited extraction‑style synthetic training data. Combine retrieval pipelines, extraction fine‑tuning, or hybrid architectures if you need precise extraction at scale.

Executive summary for decision‑makers

Soofi S is a significant, pragmatic open‑weight contribution: a Nemotron‑style MoE hybrid that prioritizes German data and long‑context throughput while publishing weights and a data inventory. For German enterprises and teams building agents or document‑heavy services, it’s an attractive base model, especially where regional control and on‑prem options matter.

That value comes with trade‑offs: weaker competition‑level math performance and a pronounced extraction failure on very long contexts unless you add extraction‑targeted training or retrieval layers. Before deploying, confirm the license, replicate throughput and extraction tests at your scale, and require safety and hallucination assessments. If you want to influence the model’s next phase, the consortium is seeking industry partners for co‑development in technical documentation, code generation and agent scenarios, an opportunity for early adopters to help close the model’s current gaps.