Skip to main content
Connic
Back to BlogIndustry Insights

Soofi S Preview: Access, Benchmarks & AI Agent Fit

Soofi S is a gated preview of Germany's 31.6B open-model project. Its benchmarks, release status, limits, and potential for self-hosted AI agents.

July 15, 2026(last updated: August 9, 2026)9 min readAuthor: Connic Research Team
Update: August 9, 2026
Status as of August 9, 2026: The available checkpoints are still gated closed-beta research previews, and the promised ungated release under a permissive license has no announced date. The preview lineup has grown: the project now lists entropy-quantized EntQuant builds of the instruct and both reasoning previews at effective 2, 3, and 4 bits per parameter. The 4-bit instruct build fits its weights in about 13 GB of GPU memory. The release, licensing, benchmark, and deployment details below reflect this current status.

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 future permissive open release. The checkpoints currently listed on Hugging Face are gated closed-beta research previews, not a general open release. 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. 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.

31.6B
total parameters
Mixture-of-experts, 128 routed experts.
~3.2B
active per token
6 experts routed per token, plus shared experts.
1M
token context
Extended in a dedicated long-context phase.
27T
training tokens
Three phases, German overweighted throughout.

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 decode speed from 4,000 out to 256,000 tokens of context: around 4,800 tokens per second per GPU at a 40,000-token context with 32 parallel requests, several times what dense models in the 14-24B range manage under the same load.

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.

Efficiency has a floor
The sparse design cuts compute per token, not memory. All 31.6B parameters have to sit in GPU memory regardless of how few are active, so the minimum hardware to serve it is that of a 30B model, even if it decodes like a much smaller one. Quantization softens this: the project lists gated GGUF, FP8, and entropy-quantized EntQuant previews for approved testers, and the 4-bit EntQuant instruct build needs about 13 GB of GPU memory for weights before the KV cache. Exact requirements still depend on serving stack, context length, and concurrency.

The benchmarks, honestly

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: an aggregate of 79.1 and the best German code aggregate in that evaluation. The technical report is unusually transparent about the rest.

GSM8K (math)86.1%
MBPP-DE (code, German)84.2%
German aggregate79.1%
HumanEval (code)73.8%
MBPP (code)70.2%
English aggregate70.1%
HumanEval-DE (code, German)65.5%
Minerva MATH-DE (math, German)56%
GPQA-Diamond (reasoning)43.4%

Selected scores from the Soofi S technical report, pass@1 for code benchmarks.

The caveats matter just as much. Against open-weight but not fully open models, Soofi S loses clearly: Qwen3.5 scores 74.6 on the English aggregate to Soofi's 70.1, with similar gaps on abstract reasoning benchmarks like GPQA. 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.

Critics have also noted that the architecture overlaps heavily with Nvidia's Nemotron 3 Nano family and argue that continual pretraining on an existing base would have cost far less than the 253,000 GPU-hours spent. That is a fair critique of the economics. It does not change the central result: a German-strong model with unusually detailed documentation, public training and evaluation code, and a largely reconstructable data recipe. Access to the current checkpoints and final licensing remain separate questions.

Open ambition, gated access today

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.

Closed beta, not an open release
The current Hugging Face model card limits checkpoint access to selected beta partners and promises a later ungated release under a permissive license. Until that release and its final terms arrive, treat any production or commercial plan as provisional. This is the single most important thing to re-check before committing to the model.

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 a compatible GGUF stack such as llama.cpp or Ollama, then register the endpoint as a custom OpenAI-compatible provider in your project settings, and reference it from the agent config:

Same agent, your endpoint
A custom provider gets a user-defined model prefix, so switching an agent from a commercial model to your own Soofi S deployment is a one-line change. For example, 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 honest caveats carry through. The base model is explicitly 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. That is why we would 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.

Run open models behind production agents

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 free

The 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: the general ungated release and final license, the instruct model graduating from preview, and the consortium's next phase, which targets agent systems directly and is looking for industry partners. If those land, the default answer to "which model do we self-host for German workloads" could change.

Frequently Asked Questions

Soofi S is a 31.6B-parameter mixture-of-experts language model from a German research consortium coordinated by the KI Bundesverband. The current checkpoints are gated closed-beta research previews, while an ungated permissive release is planned. It activates about 3.2B parameters per token, supports contexts up to 1 million tokens, and focuses on German and English.

Not as a generally available release today. The current Hugging Face model card labels Soofi S a gated closed-beta preview. Its technical report, training and evaluation code, and detailed data accounting are public, and the consortium says the final model will be released without gating under a permissive license.

In the consortium's evaluation, Soofi S leads the fully open baselines tested on German and English aggregate benchmarks, ahead of OLMo 3 32B and Apertus 70B, with especially strong German code results. Against open-weight models like Qwen3.5 it trails on English and abstract reasoning. Weak spots include German competition math, open-domain factual recall, and reliable recall over very long contexts.

It uses a hybrid architecture: only 6 of 52 layers are attention layers that keep a KV cache, and the rest combine Mamba-2 state-space layers with mixture-of-experts blocks. Decode throughput stays nearly constant from 4,000 to 256,000 tokens of context. The trade-off is memory: all 31.6B parameters must be loaded even though few are active per token.

Only if you have approved beta access. Use an instruction-tuned preview rather than the unaligned base model, serve it with a compatible stack, and connect the endpoint as an OpenAI-compatible provider. On Connic this is a custom provider in project settings, and a private inference endpoint can be reached through the bridge. Evaluate the preview on representative traffic before routing important workloads to it.

The unquantized model uses bf16 weights, so the weights alone require roughly 63 GB before runtime overhead. The project also lists gated GGUF, FP8, and entropy-quantized EntQuant previews; the 4-bit EntQuant build cuts weight memory to about 13 GB. Actual hardware requirements depend on the chosen quantization, serving stack, context length, and concurrency.

It shows promising German performance, was trained in the EU, has unusually detailed public training materials, and is designed for self-controlled infrastructure. The current preview remains gated, but the planned permissive release could become relevant to teams whose deployments depend on EU data residency and EU AI Act readiness.

More from the Blog

Industry Insights

What Is an MCP Connector? A Practical Definition

An MCP connector links an AI app to external tools and data over the Model Context Protocol. Learn how it works and when it beats a custom API integration.

July 8, 20268 min read
Industry Insights

AI Agent Platforms With EU Data Residency: 2026 Shortlist

A 2026 shortlist of AI agent platforms grouped by EU residency model, including coverage for traces, storage, and model calls.

July 6, 202612 min read
Industry Insights

Webhook vs Kafka vs SQS vs Postgres for AI Agent Triggers

Compare webhook, Kafka, SQS, and Postgres LISTEN/NOTIFY as AI agent triggers by delivery guarantees, ordering, replay, latency, and failure behavior.

June 29, 20269 min read
Industry Insights

State of AI Agents in DACH 2026

How DACH teams build, trigger, and run production AI agents in 2026: adoption, model mix, connectors, cost, reliability, and compliance, from Connic customer data.

June 27, 202612 min read
Industry Insights

Pre-built AI Agent Connectors: Platforms, Types & Checklist (2026)

Compare pre-built AI agent connector platforms, connector types, supported modes, and the delivery guarantees to verify before choosing one.

June 16, 20269 min read
Industry Insights

The Real Cost of Assembling Your Own AI Agent Stack

The real cost of assembling your own AI agent stack comes from the integration and maintenance tax between tools. Learn when buying a platform wins.

June 9, 202610 min read
Industry Insights

Best AI Agent Platforms for EU Enterprises in 2026

Ranked shortlist of AI agent platforms evaluated on EU data residency, self-hosting, MCP tool support, BYOK, EU AI Act readiness, and SLA terms. Updated July 2026.

May 19, 202616 min read
Industry Insights

AI Agent Deployment Platforms: 16 Vendors Compared (2026)

Compare 16 AI agent deployment platforms by runtime boundary, language, hosting model, connector ownership, residency, and pricing.

April 19, 202615 min read
Industry Insights

EU AI Act Enforcement: Who Investigates and What Evidence to Keep

The AI Office, national authorities, and EDPS divide EU AI Act enforcement by system and provider; teams should keep scoped governance and runtime evidence.

April 13, 202614 min read