On July 13, 2026, a German research consortium published the Soofi S technical report and began the release phase for a 31.6B-parameter language model intended for a permissive open release. The checkpoints currently listed on Hugging Face are gated closed-beta research previews, not a general open release. Fraunhofer now targets September 2026 for the base model; post-trained dialogue and agent variants follow later. In the consortium's own evaluation, Soofi S leads the fully open baselines it tested on German and English aggregates. The Soofi project is coordinated by the KI Bundesverband with Fraunhofer institutes, DFKI, universities, and companies including Ellamind and Merantix Momentum. The German Federal Ministry for Economic Affairs and Energy funded the project with roughly 20 million euros. This post looks at what is available, where the model is strong, where it is not, and what the preview could mean for production agents.
What is available today
Public materials include the technical report, training and evaluation code, and detailed data accounting. Hugging Face lists gated closed-beta checkpoints for the base model (Soofi-S-30B-A3B on Hugging Face), an instruction-tuned preview, and two reasoning previews named Isar and Rhine. Access is limited to selected beta partners while the consortium develops the final version. The Fraunhofer release page lists September 2026 as the planned base-model release; the instruction-tuned release remains undated. Training ran from late March to mid May 2026 on up to 512 Nvidia B200 GPUs in Munich, about 253,000 GPU-hours on Deutsche Telekom infrastructure, over roughly 27 trillion tokens. German content was deliberately overweighted: 7.2% of the main pretraining mix and 15.3% of the later high-quality annealing phase, compared with the roughly 5% that common training recipes reserve for all non-English data combined.
The architecture is the interesting part
Soofi S is not chasing frontier scores. It is designed to be cheap to serve. The model is a hybrid: of its 52 layers, only 6 are classic attention layers that keep a KV cache, while the rest pair Mamba-2 state-space layers with mixture-of-experts blocks. The practical consequence is that throughput barely degrades as the context grows. The consortium measured nearly constant aggregate decode throughput from 4,000 out to 256,000 tokens of context: 4,820 tokens per second per B200 GPU at a 40,000-token context with 32 parallel requests. That is a batched throughput result, not single-request generation speed.
For agent workloads this is the headline feature. Agents accumulate context: system prompts, tool schemas, tool results, retrieved documents, conversation history. A model whose serving cost stays flat as that context grows is built for exactly this shape of traffic. The consortium says as much: the stated focus for the next project phase is technical documents, code generation, and agent-based systems.
Corrected benchmark results
In the consortium's comparison, the Soofi S checkpoint scores highest among the fully open baselines tested on both German and English aggregates, ahead of OLMo 3 32B and Apertus 70B. That benchmark category describes the comparison set; it does not make the current gated Soofi checkpoint a fully open release. Its German results are the standout: a corrected aggregate of 85.3 and the best German code aggregate in that evaluation. The technical report is unusually transparent about the rest.
Selected scores from the Soofi S technical report, pass@1 for code benchmarks.
The July 22 report revision disclosed that paraphrased GPQA evaluation items had entered the training data. The authors removed GPQA-Diamond and GPQA-Diamond-DE from the tables and recomputed every model's aggregates. We have removed the earlier 43.4 GPQA score from this article; the current English and German aggregates are 77.3 and 85.3. Read the report's contamination disclosure for the root cause and remediation.
Against open-weight but not fully open models, Soofi S trails Qwen3.5: 81.3 versus 77.3 on the English aggregate and 87.6 versus 85.3 on German, with further gaps on reasoning benchmarks such as BBH and AGIEval. Competition-style math in German is weak at 56 points on Minerva MATH-DE, and open-domain factual recall runs into the limits of ~3.2B active parameters. And while long-context throughput is excellent, one RULER word-extraction subtask degrades past 32,000 tokens and drops below 5% accuracy at the longest lengths, so the 1M-token context is best treated as a throughput property, not a guarantee of reliable recall across it.
Soofi S uses the same model architecture as NVIDIA's Nemotron 3 Nano. Its differentiation is the German-English training recipe, public training and evaluation code, and detailed data documentation rather than a new architecture. Access to the current checkpoints and final licensing remain separate questions.
Soofi S remains a gated beta
The Soofi project targets a permissive open release, but the current preview does not offer permissionless public access to its checkpoints. The model card labels access as closed beta and says the final model will be released openly without gating. About 99% of the training mixture is publicly obtainable and can be reconstructed where source licenses permit; the remainder includes commercially licensed German newspaper archive content that cannot be redistributed. The training and evaluation code and detailed data accounting are public, but the current model weights and final license terms are not generally available.
Can you run Soofi S with Ollama?
Yes, but only with approved beta access as of September 1, 2026. The Soofi team publishes gated GGUF previews for Instruct, Isar, and Rhine and an official local hosting setup that imports them into Ollama. As of September 1, the project does not publish an ungated model through the Ollama library. Soofi documents a token redirect issue when direct ollama pull hf.co/... uses gated authentication, so its setup downloads the file with the Hugging Face CLI and creates the local Ollama model from the supplied Modelfile.
| Format | Published size | What that means |
|---|---|---|
| GGUF Q5_K_M | 26 GB | The quantization used by the official Ollama setup. Runtime and context need additional memory. |
| GGUF Q8_0 | 33.6 GB | Larger GGUF option; no official full-GPU minimum is published. |
| Unquantized GGUF | 63.2 GB | Weights alone; serving overhead is additional. |
| EntQuant 4-bit | About 13 GB of GPU memory for weights | A separate vLLM format that does not load in Ollama or llama.cpp. |
The official llama.cpp profile defaults to eight GPU-offloaded layers, documented as using roughly 8 GB of VRAM on a 12 GB card. That is a llama.cpp partial-offload configuration, not an Ollama minimum. The repository does not publish a full-offload minimum or a comparative speed benchmark. Size memory for the model file, runtime overhead, context, and concurrency.
From the root of the official hosting repository, start the Ollama profile and run the documented Instruct import below. These commands were verified against the Soofi repository on September 1, 2026; we could not execute the gated 26 GB download without approved access.
docker compose up -d
pip install -U "huggingface_hub[cli]"
export HF_TOKEN=hf_xxx
hf download Soofi-Project/Soofi-S-Instruct-Preview-GGUF \
soofi-s-instruct-preview-Q5_K_M.gguf \
--local-dir files/
docker compose exec ollama ollama create \
soofi-s-instruct-preview \
-f /files/soofi-s-instruct-preview.modelfile
docker compose exec ollama ollama run soofi-s-instruct-previewOllama is a practical local test path, but the Soofi repository warns that Ollama's Go templates do not fully reproduce the model's embedded Jinja template, reasoning history, or native tool-calling format. The Instruct preview can also place its answer in the reasoning channel unless the API request sends think: false. For agent tool calling, the official repository documents its llama.cpp server path with Jinja enabled as the option that preserves the embedded chat template.
Running Soofi S behind an agent
Connic keeps BYOK and custom OpenAI-compatible endpoints as first-class options alongside its managed model catalog. If you have approved beta access, serve the instruction-tuned preview with vLLM or the official llama.cpp path when native tool calling matters, then register the endpoint as a custom OpenAI-compatible provider in your project settings, and reference it from the agent config:
model: vllm/soofi-s-instruct uses a Connic provider alias; it is not an official Soofi model ID. If the inference box lives inside your private network, agents reach it through the bridge instead of the public internet. Guardrails, judges, traces, and approvals work the same regardless of which model answers.The operational limits remain: the base model is not meant for direct use without post-training, the instruct and reasoning variants are gated previews, and we found no public commercial SLA or support offering. Beta partners may have different arrangements, but teams evaluating a self-hosted preview should plan to own patches, security updates, and incident response. Put Soofi S on a slice of real traffic behind an A/B test, with judges scoring its output, before routing anything important to it. Our DACH report found 23% of production agents in the region already run open-weight or self-hosted models, so the operational pattern is well established; what has been missing is a German-first model worth the effort.
Connect a self-hosted Soofi S endpoint as a custom provider and put guardrails, judges, and traces around it, alongside whichever commercial models you already use.
Get started freeThe sovereignty question
Soofi S will be discussed mostly in sovereignty terms, and it is worth being precise about what it does and does not settle. It does not make the strongest closed or open-weight models redundant; the benchmark gaps are real. What it provides today is evidence that a German-strong sovereign model with efficient long-context serving is technically plausible, plus a detailed report and public training and evaluation code. Approved beta testers can run preview checkpoints entirely on infrastructure they control; general access, final terms, and production maturity remain outstanding. For teams whose deployments hinge on EU AI Act readiness and keeping data inside the EU, that combination is the point, independent of leaderboard positions.
Three things to watch from here: whether the planned September base release ships with its final license, when the instruction-tuned model becomes generally available, and how its tool calling behaves outside the beta. If those land, the default answer to "which model do we self-host for German workloads" could change.