Germany's Soofi consortium has trained a German-English foundation model on domestic cloud infrastructure, but its public checkpoint, benchmark scope and unresolved data-contamination allegation leave the project's openness and production economics unproven.
Soofi S is a serious German infrastructure and model-development effort, not yet the open industrial product its launch language implies. The project has shown that a German consortium can train a large model on domestic cloud capacity. It has not yet established a clean benchmark lead, an ungated release or lower real-world deployment cost.
The Soofi team describes Soofi S 30B-A3B as a hybrid Mamba-Transformer mixture-of-experts base model for German and English. Its paper says the model was pretrained on roughly 27 trillion tokens and activates about 3 billion of its 30 billion parameters for each token. That sparse design is intended to reduce inference work while limiting the attention cache that grows with long inputs.
The word “open,” however, describes a promised end state more accurately than the artifact currently offered. The model card calls the checkpoint a closed-beta research artifact for selected partners and says it is not an open release. The listed custom license is unfinished, and the base model has not been instruction-tuned, aligned or safety-tuned. It is designed for continued pretraining, fine-tuning and research, not direct use as an assistant.
The project’s own public page likewise says no general release for direct use has occurred and asks organizations to apply for the next industrial test phase. Its pretraining repository contains scripts, configurations and documentation but said open model weights were “coming soon” when archived.
The public materials promise an ungated, permissively licensed final model, selected intermediate checkpoints, per-source data accounting, hyperparameters, and training and evaluation code. That is an unusually broad disclosure plan. It remains a release commitment, not a description of what any organization can download and deploy today.
The public specifications also need reconciliation. The paper reports roughly 27 trillion tokens and 3 billion active parameters; the model card reports about 25 trillion tokens and approximately 3.5 billion active parameters. Those figures could refer to different checkpoints or accounting conventions, but the retained documents do not explain the difference.
Soofi’s technical contribution is narrower than the phrase “European model” can suggest. The project repository says its complete training code produced a 30-billion-parameter model based on Nvidia’s Nemotron 3 Nano architecture. A technical comparison of the project results says Soofi adopted that architecture without modification.
Nvidia describes Nemotron 3 Nano as a 31.6-billion-parameter hybrid model with 3.2 billion active parameters. Nvidia released base and post-trained weights, its training recipe, and data for which it holds redistribution rights. That existing model is both the enabling foundation and the clearest substitute against which Soofi’s incremental value should be measured.
The German work is concentrated in the data mixture, the decision to pretrain from scratch, reproducibility, local operations and industrial testing. The technical comparison says German made up 7.2% of Soofi’s first phase and 15.3% of its second, versus about 5% for all non-English languages in Nvidia’s reference recipe. The mixture included German web and reference material, machine-translated and synthetic text, and a commercially licensed corpus of 193 million newspaper articles from 916 publications.
That commercial corpus accounts for 1.3% of the mixture, according to the analysis, preventing every training token from being freely redistributed. The team nevertheless says about 99% of the mixture can be reconstructed and argues that the planned release meets the Open Source AI Definition. The unfinished license and gated checkpoint mean that claim cannot yet be evaluated against the final package.
Soofi is a consortium rather than a conventional venture-backed model company with a founder story. It is coordinated by the German AI Association and brings together Fraunhofer institutes, DFKI, several universities, and the companies Ellamind and Merantix Momentum. Its launch announcement says the project is funded by Germany’s Federal Ministry for Economic Affairs and Energy under the IPCEI-CIS/8ra initiative, with financing from the European Union’s NextGenerationEU program. That structure makes public auditability, not only model performance, part of the project’s credibility test.
The team’s paper claims the highest aggregate English and German evaluation scores among the fully open models it tested, ahead of OLMo 3 32B and Apertus 70B, plus the best English and German code aggregates among 17 open base models. This is a defined peer comparison, not evidence that Soofi beats every open-weight model or a frontier commercial system.
The distinction matters. Michael Fromm, Soofi’s head of pretraining data, said Qwen3.5 was ahead in raw capability and framed Soofi’s proposed edge as capability combined with throughput and openness, according to reporting on the release. The same report quotes German lawmaker Karl Lauterbach saying Soofi does not compete with Claude or OpenAI’s top models. Those are different comparison groups and different procurement choices.
Within the team’s suite, the reported profile is uneven. The technical analysis gives Soofi scores of 73.8% on HumanEval, 70.2% on MBPP and 84.2% on a German MBPP variant. It tied Qwen3.5 35B-A3B at 61.2 on INCLUDE-DE, a regional-knowledge test. But it trailed Qwen3.5 and Gemma 3 27B on German competition math and lagged on NaturalQuestions factual retrieval.
One evaluation input now faces a more fundamental challenge. In a public critique, model developer Daryoush Daniel Vaziri alleged that the AIML-TUDA/QA-base training dataset contained all 198 GPQA-Diamond questions with answer labels. He said Soofi used QA-base for 10 epochs and that GPQA produced the report’s largest single capability-index gain, 9.6 points. Vaziri also alleged overlap with BLiMP, TruthfulQA and Inverse Scaling test material.
Those are attributed allegations, not an independent audit. The archived Soofi paper, project pages and model card do not answer them, and the retained body of Vaziri’s post does not include a Soofi reply. The defensible conclusion is therefore narrower than the provisional launch coverage and narrower than a declaration of confirmed contamination: the GPQA result should not carry evidentiary weight until the consortium publishes a response, a decontamination method and revised aggregates.
Vaziri further estimated that roughly 78% of effective training tokens came from Nemotron sources and contrasted 253,000 B200 GPU-hours of from-scratch training with a 1.8-point English aggregate improvement over Nemotron. That calculation is also a critic’s claim. It nevertheless identifies the correct economic question: whether retraining an inherited design creates enough German-language and control value to justify the compute and public support.
The efficiency evidence is promising but workload-specific. In the project’s reported tests, Soofi generated about eight times as many tokens per second per GPU as dense models in the 14-billion-to-24-billion-parameter range at a 40,000-token context with 32 concurrent requests. Throughput stayed nearly flat from 4,000 to 256,000 tokens. Qwen3.5 35B-A3B, another hybrid model, showed similar behavior.
The same evaluation found a sharp quality boundary. On a RULER task requiring extraction of frequently occurring words, Soofi’s hit rate fell to about 3% beyond 32,000 tokens; the comparable Nemotron model achieved 60% to 64%. The authors attributed that deficit to long-document training data that lacked synthetic extraction tasks. A near-constant cache and high token throughput therefore show that the model can process long prompts efficiently, not that it reliably retrieves information from all of them.
The retained sources provide no production pricing, measured cost per successful task, service-level commitment or comparable energy figure. The consortium’s claim of lower energy consumption and the reported throughput tests do not fill that gap. Buyers would still need workload-specific measurements that include hardware utilization, latency, accuracy and operational support.
The training run used up to 512 Nvidia B200 GPUs and about 253,000 GPU-hours on Deutsche Telekom’s Industrial AI Cloud in Munich, according to the technical analysis. Telekom had earlier announced an infrastructure contract in the tens of millions of euros from Leibniz University Hannover. It said more than 1,000 GPUs in roughly 130 DGX B200 systems would be dedicated to Soofi within a 10,000-GPU facility.
That earlier announcement described a planned model of roughly 100 billion parameters intended to succeed Teuken7B. Soofi S, at about 30 billion total parameters, is the first component of a proposed family rather than fulfillment of that larger specification. The contract still represents meaningful domestic compute capacity and gives the consortium more control over operations and data location.
It does not make the stack independent of foreign technology. Soofi relies on Nvidia accelerators, Nvidia systems and an Nvidia-originated architecture. “Sovereign” here means that training and governance are located in Germany and the consortium intends to make the model inspectable and adaptable. It does not mean a wholly European hardware or model-design supply chain.
Four pieces of evidence would resolve the gap between the project’s strategic promise and the product available now:
Until then, Soofi S’s strongest verified result is institutional: Germany assembled public research, commercial cloud capacity and a disclosed training process at substantial scale. Whether that stack produces a better industrial choice remains a question for an audited evaluation, the final license and production evidence—not the launch ranking alone.
Shanghai AI Laboratory has released Intern-S2-Preview-397B as downloadable weights and a hosted API, but its own 35B alternative offers the same context window and a far smaller disclosed scale while the larger model still lacks auditable comparative scores, pricing and physical-validation results.
Mixfont's free Decoy Font gives each glyph a sharp decoy and a blurred intended letter, creating a hurdle for pixel-based readers while leaving selectable text, known-image techniques and accessibility costs outside its protection.
Zhipu reportedly reached $1 billion in annual recurring revenue in July, roughly four times a March estimate, but the unconfirmed run rate is not annual sales and still sits far ahead of recognized cloud revenue while margins remain thin.
Anthropic’s 20-year lease gives TeraWulf a customer for 401 megawatts of future AI infrastructure, but rent starts only after delivery and the proposed utility deal passes power and grid costs to TeraWulf, leaving financing and project margin unresolved.
OpenAI has put conversations and Projects back in its redesigned ChatGPT desktop app and enabled cloud Work threads to move across devices, correcting the launch's biggest usability failures without merging local Work or Codex histories.
Twenty-nine countries signed an agreement creating WAICO as an independent intergovernmental organization, while China paired the launch with capacity-building offers that are not yet confirmed as WAICO programs.
Meta reportedly plans to put departing AWS compute executive Dave Brown to work on its data-center buildout, adding hyperscale operating experience while leaving any customer-facing cloud business conditional and undefined.
Moonshot AI has made Kimi K3 available through its apps and API, pairing a 2.8-trillion-parameter architecture with early frontier-level results, but the model's open-weight claim cannot be tested until its weights and technical report arrive.
CIA Director John Ratcliffe said US intelligence is consistent with an estimate that Russian recruits last 20 to 30 minutes on Ukraine’s battlefield, but the public trail leads to an unsourced claim about assault troops and does not establish a representative average.
China has paired a five-year AI training offer for developing countries with cooperation centers, a weather-warning rollout and a new 29-country organization. The package gives Beijing a platform for influence, but no budget, selection rules or delivery timetable has been published.
Federal intervention changed the launches of Anthropic's Claude Fable 5, Claude Mythos 5 and OpenAI's GPT-5.6, but through different legal and voluntary channels. The result is real government leverage over early customers without a common security threshold or durable review process.
A proposal developed with Treasury Secretary Scott Bessent would put an independent AI regulator under SEC oversight, but President Trump had not reviewed it and its tests, funding and enforcement authority remained unresolved.
San Francisco has demanded that Apple and Google cut off 13 apps capable of producing nonconsensual sexual deepfakes. The threatened case could test California’s new liability rules, but the confidential app list and the law’s “primary purpose” definition leave a central question unresolved.
Anthropic will keep Claude Fable 5 inside Max and Team Premium plans from July 20 at 50% of their usage limits, while Pro and Team Standard customers move to separately billed credits. The durable change is who gets bundled access, not a larger allowance for top-tier users.
HKT plans a 3.2 Tbps data-centre route from Lok Ma Chau to Tseung Kwan O by the end of 2026, but has not disclosed customer bandwidth, price, project cost or end-to-end latency, while Sandy Ridge's developer expects to start within 42 months of its land award.
Apple briefly topped Nvidia during Friday trading before Nvidia closed about $6 billion ahead. The filings show why that ranking is a weak AI verdict: Apple’s operating momentum predates its new Siri, while Nvidia’s rapid growth comes with concentrated customers and large forward commitments.
Anthropic has proposed paying Meta as much as $10 billion over two years for AI computing capacity, but early exit rights, undisclosed capacity and Meta's own reliance on outside infrastructure leave the economics—and the case for a durable Meta cloud business—unresolved.
Barracuda says it detected more than one million retail-themed phishing emails using hidden filler text since April, but its published research does not show how many passed security controls or whether AI-generated filler performs better than older variants.
A federal judge denied emergency relief to 26 Meta workers challenging allegedly AI-assisted layoff selections, finding no near-term irreparable harm while leaving the discrimination claims, a preliminary-injunction request and the role of Meta's internal tools unresolved.
Google says it fixed a Gemini flaw that let someone with physical access to a locked Android phone send messages and reconnect apps without the expected PIN check, but it has not identified the affected devices, delivery channel, or software version that would let users verify the repair.