{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-07-28T05:20:57.982Z","headline":"德国AI协会发布开源模型Soofi S，在英语和德语基准测试中领先","description":"德国AI协会协调的研究联盟发布开源大语言模型Soofi S 30B-A3B。该模型总参数量316亿，每个token仅激活约32亿参数，采用Mamba-2与标准注意力层混合的MoE架构。模型完全在德国电信慕尼黑工业AI云上训练，训练数据中德语占比从第一阶段的7.2%提升至第二阶段的15.3%。在基准测试中，Soofi S在所有完全开源模型中取得英语和德语综合最高分，超越OLMo 3 32B和Apertus 70B。在HumanEval上得分73.8%，MBPP得分70.2，德语版MBPP得分84.2。上下文窗口支持最高100万token，在4万token长度下，生成吞吐量约为同规模稠密模型的8倍。模型权重已开源。","url":"https://www.aioga.com/news/cmrj6actv0651bilkm5pfz6ub/","mainEntityOfPage":"https://www.aioga.com/news/cmrj6actv0651bilkm5pfz6ub/","datePublished":"2026-07-13T11:41:01.000Z","dateModified":"2026-07-13T11:41:01.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://the-decoder.com/german-ai-consortium-releases-soofi-s-an-open-30b-model-that-tops-benchmarks-in-both-english-and-german","https://aihot.virxact.com/items/cmrj6actv0651bilkm5pfz6ub"],"canonicalUrl":"https://www.aioga.com/news/cmrj6actv0651bilkm5pfz6ub/","directAnswer":{"@type":"Answer","text":"德国AI协会协调的研究联盟发布Soofi S 30B-A3B。模型共316亿参数，每个生成token约激活32亿参数，采用结合Mamba-2层与标准注意力层的混合专家架构，模型权重已经开放。","url":"https://www.aioga.com/news/cmrj6actv0651bilkm5pfz6ub/","dateCreated":"2026-07-13T11:41:01.000Z","author":{"@type":"Organization","@id":"https://www.aioga.com/authors/aioga-editorial/#editorial-team","name":"Aioga Editorial Team","url":"https://www.aioga.com/authors/aioga-editorial/"}},"evidence":[{"@type":"CreativeWork","name":"the-decoder.com source article","url":"https://the-decoder.com/german-ai-consortium-releases-soofi-s-an-open-30b-model-that-tops-benchmarks-in-both-english-and-german","datePublished":"2026-07-13T11:41:01.000Z","provider":{"@type":"Organization","name":"the-decoder.com","url":"https://the-decoder.com/german-ai-consortium-releases-soofi-s-an-open-30b-model-that-tops-benchmarks-in-both-english-and-german"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.virxact.com/items/cmrj6actv0651bilkm5pfz6ub","datePublished":"2026-07-13T11:41:01.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.virxact.com/items/cmrj6actv0651bilkm5pfz6ub"}}],"aggregationSource":"The Decoder：AI News（RSS）","originalPublisher":{"name":"the-decoder.com","url":"https://the-decoder.com/german-ai-consortium-releases-soofi-s-an-open-30b-model-that-tops-benchmarks-in-both-english-and-german"},"article":{"id":"cmrj6actv0651bilkm5pfz6ub","slug":"cmrj6actv0651bilkm5pfz6ub","url":"https://www.aioga.com/news/cmrj6actv0651bilkm5pfz6ub/","title":"德国AI协会发布开源模型Soofi S，在英语和德语基准测试中领先","title_en":"German AI consortium releases Soofi S， an open 30B model that tops benchmarks in both English and German","summary":"德国AI协会协调的研究联盟发布开源大语言模型Soofi S 30B-A3B。该模型总参数量316亿，每个token仅激活约32亿参数，采用Mamba-2与标准注意力层混合的MoE架构。模型完全在德国电信慕尼黑工业AI云上训练，训练数据中德语占比从第一阶段的7.2%提升至第二阶段的15.3%。在基准测试中，Soofi S在所有完全开源模型中取得英语和德语综合最高分，超越OLMo 3 32B和Apertus 70B。在HumanEval上得分73.8%，MBPP得分70.2，德语版MBPP得分84.2。上下文窗口支持最高100万token，在4万token长度下，生成吞吐量约为同规模稠密模型的8倍。模型权重已开源。","source":"The Decoder：AI News（RSS）","sourceUrl":"https://the-decoder.com/german-ai-consortium-releases-soofi-s-an-open-30b-model-that-tops-benchmarks-in-both-english-and-german","aiHotUrl":"https://aihot.virxact.com/items/cmrj6actv0651bilkm5pfz6ub","publishedAt":"2026-07-13T11:41:01.000Z","category":"模型更新","score":70,"selected":true,"articleBody":["Soofi S is one of the first large language models trained entirely on Deutsche Telekom's Industrial AI Cloud in Munich. The open 30B model uses a lean hybrid architecture and a training mix deliberately weighted toward German.","Soofi S is a mixture-of-experts model. It contains 31.6 billion parameters in total but activates only about 3.2 billion per generated token. That puts its compute cost closer to a 3B model than a conventional 30B model. The consortium adopts the architecture of Nvidia's Nemotron 3 Nano：https://the-decoder.com/nvidias-nemotron-3-ultra-becomes-the-smartest-open-us-model-but-china-still-leads/ without modification, a hybrid design combining Mamba-2 layers with standard attention layers. Ad","The key difference from typical transformers is memory behavior. In conventional models, the KV cache that stores previous tokens for attention computation grows linearly with context length. With long inputs and many parallel requests, reloading that cache becomes a bottleneck. Only 6 of Soofi S's 52 layers maintain such a cache at all. Ad DEC_D_Incontent-1","The practical payoff shows up in generation throughput. At a context length of 40,000 tokens with 32 parallel requests, Soofi S generates roughly eight times more tokens per second per GPU than dense models in the 14 to 24 billion parameter range. While throughput drops significantly for conventional models as context grows, Soofi S stays nearly flat from 4,000 to 256,000 tokens. The only model that shows similar behavior in the measurements is Alibaba's Qwen3.5 35B-A3B：https://the-decoder.com/alibabas-open-qwen-3-5-takes-aim-at-gpt-5-mini-and-claude-sonnet-4-5-at-a-fraction-of-the-cost/, which also uses a hybrid architecture.","The consortium processed about 27 trillion tokens in total, split across three phases. In the first phase, the model learns language fundamentals from roughly 20 trillion tokens drawn from a broad mix of web, code, math, and domain-specific texts. A second phase follows with about 6 trillion tokens from higher-quality sources, designed to sharpen the patterns learned earlier. A shorter third phase then extends the context window by training on very long documents of up to one million tokens. Ad","The deliberate focus on German is central. In the first phase, German makes up 7.2 percent of the training mix; in the second phase, that share rises to 15.3 percent. In Nvidia's Nemotron reference recipe, all non-English languages combined account for only about 5 percent.","For data sources, the consortium combines German web text from HPLT, the openly licensed German Commons corpus, German portions of FinePDFs and FineWiki, and the commercially licensed Genios corpus containing 193 million newspaper articles from 916 German publications. Machine-translated and synthetically generated German texts round out the mix. Ad DEC_D_Incontent-2","In evaluations against 16 other open models, Soofi S leads all fully open models on aggregate scores for both German and English, according to the report. That includes OLMo 3 32B from the Allen Institute for AI and Apertus 70B from ETH Zurich and EPFL. Against every European sovereign baseline, the model comes out ahead on all German benchmarks in the suite, sometimes by double-digit margins. Ad","On code benchmarks, Soofi S scores 73.8 percent on HumanEval, 70.2 on MBPP, and 84.2 on the German MBPP variant, the best results among open-source peers. On INCLUDE-DE, a test for Germany-specific regional knowledge, Soofi S ties for first place at 61.2 points with the larger Qwen3.5 35B-A3B. Compared to the Nemotron baseline, the German data recipe improves language proficiency by 15.1 points and the science test GPQA-Diamond by 9.6 points, without sacrificing English performance.","Soofi S doesn't do as well on German competition math, where it scores 56 points on Minerva MATH-DE, well behind Qwen3.5 35B-A3B (76.5) and Gemma 3 27B：https://the-decoder.com/google-releases-new-gemma-3-open-model-family/ (65.6). It also lags on open factual retrieval in NaturalQuestions. The latter likely relates to having only 3 billion active parameters, which can store less world knowledge：https://the-decoder.com/sinas-open-model-vibethinker-3b-aims-to-show-reasoning-compresses-well-but-factual-knowledge-doesnt/ than a dense 27B model.","The RULER long-context test also reveals a specific weakness: When the model has to extract frequently occurring words from a long text, Soofi S's hit rate drops to around 3 percent beyond 32,000 tokens of context, while the comparable Nemotron model still manages 60 to 64 percent. The authors attribute this to the fact that their long-context training data contains many long documents but lacks synthetic data designed for extraction tasks. On the remaining twelve RULER tasks, both models perform about the same.","The training run took place between March and May on up to 512 Nvidia B200 GPUs at Deutsche Telekom's Industrial AI Cloud：https://the-decoder.com/10000-nvidia-blackwell-gpus-set-to-increase-germanys-ai-capacity-by-50-percent/ in Munich, totaling about 253,000 GPU-hours. According to the report, the facility runs entirely on renewable energy, is cooled with water from the Eisbach canal, and feeds waste heat into the surrounding Tucherpark neighborhood. Soofi S was one of the first major training runs on this infrastructure.","Behind Soofi is a consortium of German research institutions and companies, coordinated by the German AI Association and funded by the German Federal Ministry for Economic Affairs and Energy as part of the European IPCEI-CIS program.","Participants include the Fraunhofer Institutes IAIS and IIS, the German Research Center for Artificial Intelligence (DFKI), TU Darmstadt, the University of Würzburg, the L3S Research Center, the Berlin University of Applied Sciences, and AI companies Ellamind and Merantix Momentum. The project's goal is to build an open European AI model family that can run on sovereign infrastructure and be tested in industrial applications.","The researchers are releasing model weights along with selected intermediate checkpoints：https://huggingface.co/collections/Soofi-Project/soofi-s-beta-models, the complete training and evaluation code, and a detailed data inventory listing raw token counts, epoch numbers, and effective contributions per source. Sources that were reviewed but excluded are also documented. According to the team, this means Soofi S meets the Open Source AI Definition 1.0 from the Open Source Initiative：https://the-decoder.com/open-source-initiative-releases-first-formal-definition-of-open-source-ai/.","A stricter proposal for a European open-data definition, which would require every single training token to be freely distributable, isn't met because of the 1.3 percent share of Genios data, which carries a commercial license. The report says about 99 percent of the training mix can be independently reconstructed. The exact license for the model's release hasn't been finalized yet.","As lead author Michael Fromm writes：https://x.com/effi288/status/2075904321707798699, Soofi S positions itself between broadly multilingual European sovereignty projects like EuroLLM or Teuken：https://the-decoder.com/eu-project-releases-7b-model-that-speaks-24-european-languages/, which cover many languages, and the highest-performing international open-weight models. According to the project website, the consortium is looking for industry partners for the next phase to test the model in applications involving technical documents, code generation, and agent-based systems.","Stay in the loop on AI. Clear, useful, no fluff.","Follow The Decoder for AI news, background stories and expert analyses.","The Decoder：https://the-decoder.com/"],"articleImages":[{"sourceUrl":"https://the-decoder.com/wp-content/uploads/2026/07/soofi-s-03-data-mixture.jpg","alt":"Flow chart showing the training data mix across three phases. Seven categories including English Web, Code, Reasoning, Math, and German shift from Phase 1 (about 23T tokens) through Phase 2 (about 6T) to Phase 3 (about 188B), with the German share rising from 7.2 to 15.3 percent.","afterParagraph":4,"url":"/media/articles/cmrj6actv0651bilkm5pfz6ub/bc1678a2d239b8cb.jpg"}],"mediaStatus":"ok","articleBodyZh":["Soofi S是首批完全在慕尼黑德国电信工业人工智能云上训练的大型语言模型之一。OPEN 30B模型采用精益混合架构和故意向德国加权的训练组合。","Soofi S是一个专家混合模型。它总共包含316亿个参数，但每个生成的代币仅激活约32亿个。这使得它的计算成本比传统的30B模型更接近3B模型。 联合体采用英伟达Nemotron 3 Nano架构： https://the-decoder.com/nvidias-nemotron-3-ultra-becomes-the-smartest-open-us-model-but-china-still-leads/未经修改，采用Mamba-2层与标准关注层相结合的混合设计。广告","与典型变压器的关键区别在于记忆行为。在传统模型中，存储用于注意力计算的先前令牌的KV缓存随上下文长度线性增长。对于长输入和许多并行请求，重新加载缓存成为瓶颈。Soofi S的52层中只有6层保留了这样的缓存。广告DEC_D_Incontent-1","实际回报体现在发电吞吐量上。Soofi S的上下文长度为40,000个令牌， 32个并行请求，每个GPU每秒生成的令牌大约是140亿到240亿个参数范围内密集模型的8倍。虽然随着环境的增长，传统模型的吞吐量显着下降，但Soofi S的吞吐量几乎保持不变，从4,000个代币降至256,000个代币。 在测量中显示类似行为的唯一模型是阿里巴巴的Qwen3.5 35B-A3B： https://the-decoder.com/alibabas-open-qwen-3-5-takes-aim-at-gpt-5-mini-and-claude-sonnet-4-5-at-a-fraction-of-the-cost/，它也使用混合架构。","该联盟总共处理了大约27万亿个代币，分为三个阶段。在第一阶段，该模型从大约20万亿个令牌中学习语言基础知识，这些令牌来自网络、代码、数学和特定领域的广泛混合文本。第二阶段接下来是来自更高质量来源的大约6万亿个代币，旨在锐化之前学习的模式。 然后，较短的第三阶段通过对多达一百万个令牌的非常长的文档进行培训来扩展上下文窗口。广告","对德语的刻意关注是核心。在第一阶段，德语占培训组合的7.2%；在第二阶段，这一比例上升至15.3%。在Nvidia的Nemotron参考配方中，所有非英语语言合计仅占约5%。","对于数据源，该联盟结合了来自HPLT的德语Web文本、公开许可的德国公共语料库、FinePDF和FineWiki的德语部分以及商业许可的Genios语料库，其中包含来自916家德国出版物的1.93亿篇报纸文章。机器翻译和合成生成的德语文本完美融合。广告DEC_D_Incontent-2","根据该报告，在对其他16个开放模型的评估中， Soofi S在德语和英语的总分上领先于所有完全开放的模型。其中包括来自艾伦人工智能研究所的OLMo 3 32B和来自苏黎世联邦理工学院和EPFL的Apertus 70B。相对于每个欧洲主权基准，该模型在套件中的所有德国基准上都领先，有时以两位数的利润率领先。广告","在代码基准测试中， Soofi S在HumanEval上的得分为73.8%，在MBPP上的得分为70.2，在德国MBPP变体上的得分为84.2，这是开源同行中最好的结果。在针对德国特定区域知识的INCLUDE-DE测试中， Soofi S以61.2分与较大的Qwen3.5 35B-A3B并列第一。 与Nemotron基线相比，德国数据配方将语言熟练度提高了15.1分，科学测试GPQA-Diamond提高了9.6分，而不会牺牲英语表现。","Soofi S在德国竞争数学方面表现不佳，在Minerva MATH-DE上得分56分，远远落后于Qwen3.5 35B-A3B （ 76.5 ）和Gemma 3 27B： https://the-decoder.com/google-releases-new-gemma-3-open-model-family/ （ 65.6 ）。它在NaturalQuestions中的开放式事实检索方面也滞后。 后者可能与只有30亿个活动参数有关，与密集的27B模型相比，它可以存储更少的世界知识： https://the-decoder.com/sinas-open-model-vibethinker-3b-aims-to-show-reasoning-compresses-well-but-factual-knowledge-doesnt/。","标尺长上下文测试还揭示了一个特定的弱点：当模型必须从长文本中提取频繁出现的单词时， Soofi S的命中率在32,000个上下文令牌之外下降到约3 ％，而可比较的Nemotron模型仍然管理着60 ％至64 ％。 作者将此归因于他们的长上下文训练数据包含许多长文档，但缺乏专为提取任务设计的合成数据。在其余的十二个标尺任务中，两个模型的执行情况大致相同。","培训于3月至5月期间在德国电信位于慕尼黑的工业AI云（ https://the-decoder.com/10000-nvidia-blackwell-gpus-set-to-increase-germanys-ai-capacity-by-50-percent/ ）上使用多达512个Nvidia B200 GPU进行，总计约253,000个GPU小时。 根据该报告，该设施完全依靠可再生能源运行，用Eisbach运河的水冷却，并向周围的Tucherpark社区供给废热。Soofi S是这一基础设施的首批主要培训项目之一。","Soofi背后是一个由德国研究机构和公司组成的联盟，由德国人工智能协会协调，由德国联邦经济事务和能源部资助，作为欧洲IPCEI-CIS计划的一部分。","参与者包括弗劳恩霍夫研究所IAIS和IIS、德国人工智能研究中心（ DFKI ）、达姆施塔特工业大学、维尔茨堡大学、L3S研究中心、柏林应用科学大学以及人工智能公司Ellamind和Merantix Momentum。该项目的目标是建立一个开放的欧洲人工智能模型家族，可以在主权基础设施上运行，并在工业应用中进行测试。","研究人员正在发布模型权重以及选定的中间检查点： https://huggingface.co/collections/Soofi-Project/soofi-s-beta-models、完整的训练和评估代码以及详细的数据清单，其中列出了原始令牌计数、纪元编号和每个来源的有效贡献。还记录了已审核但排除的来源。 根据该团队的说法，这意味着Soofi S符合开源倡议（ Open Source Initiative ）的开源AI定义1.0： https://the-decoder.com/open-source-initiative-releases-first-formal-definition-of-open-source-ai/。","一项更严格的欧洲开放数据定义提案要求每一个培训令牌都是可自由分发的，但由于Genios数据的1.3%份额（带有商业许可证）而没有得到满足。报告称，大约99%的训练组合可以独立重建。该模型发布的确切许可证尚未最终确定。","正如主要作者Michael Fromm所写： https://x.com/effi288/status/2075904321707798699， Soofi S将自己定位在广泛的多语言欧洲主权项目之间，如EuroLLM或Teuken： https://the-decoder.com/eu-project-releases-7b-model-that-speaks-24-european-languages/，涵盖多种语言，以及性能最高的国际开放权重模型。 根据该项目网站，该联盟正在寻找下一阶段的行业合作伙伴，以便在涉及技术文档、代码生成和基于代理的系统的应用程序中测试该模型。","随时掌握人工智能的最新动态。清晰、实用、无绒毛。","关注解码器，了解AI新闻、背景故事和专家分析。","解码器： https://the-decoder.com/"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"德国AI协会协调的研究联盟发布Soofi S 30B-A3B。模型共316亿参数，每个生成token约激活32亿参数，采用结合Mamba-2层与标准注意力层的混合专家架构，模型权重已经开放。","background":"Soofi S完全在德国电信位于慕尼黑的工业AI云上训练。训练分为三个阶段，共处理约27万亿token；德语数据占比由第一阶段的7.2%提升至第二阶段的15.3%，第三阶段使用最长100万token的文档扩展上下文。","viewpoint":"Aioga判断，Soofi S的主要看点是以较低的单token激活参数量兼顾德语与英语评测，并通过混合架构改善长上下文吞吐表现。其优势目前来自联盟报告和特定测试设置，仍需更多独立评测验证。","implications":"在报告所列16个开放模型对比中，Soofi S的德语和英语综合成绩领先完全开放模型，包括OLMo 3 32B与Apertus 70B。值得关注的是，这可能为重视德语数据和本地训练基础设施的模型研发提供参考。","nextStep":"后续应关注模型权重、训练与评测材料的开放范围，并在一致硬件、并发数和上下文长度下复测吞吐量。同时可核验HumanEval、MBPP及德语版MBPP成绩，观察独立测试能否复现报告结论。","evidenceRefs":["title","summary","articleBody","source"],"status":"published","aiGenerated":true,"autoApproved":true,"generatedBy":"aioga-editorial:gpt-5.6-sol","reviewedBy":"aioga-editorial-review:gpt-5.6-sol","generatedAt":"2026-07-26T16:49:36.723Z","sourceHash":"4aa0c951b9a5a8e3","review":{"approved":true,"groundedness":95,"clarity":91,"duplicationRisk":12,"blockingIssues":[],"notes":["“改善长上下文吞吐表现”是基于来源测试结果的概括，来源明确限定了上下文长度、并发请求数和对比模型范围；正式发布时可保留这些测试条件以避免泛化。","“这可能为重视德语数据和本地训练基础设施的模型研发提供参考”属于合理推论，已使用“可能”限定，不构成将观点冒充事实。","“完全开放模型”应沿用来源中的 fully open 定义，必要时可进一步说明其与一般“开放模型”的区别。"]},"validation":{"passed":true,"mode":"ai-auto","revisions":0,"checks":["schema","length","source-attribution","low-source-overlap","no-html","independent-ai-review"]}},"tags":["模型更新","The Decoder：AI News（RSS）"],"translations":{"zh-CN":{"title":"德国AI协会发布开源模型Soofi S，在英语和德语基准测试中领先","summary":"德国AI协会协调的研究联盟发布开源大语言模型Soofi S 30B-A3B。该模型总参数量316亿，每个token仅激活约32亿参数，采用Mamba-2与标准注意力层混合的MoE架构。模型完全在德国电信慕尼黑工业AI云上训练，训练数据中德语占比从第一阶段的7.2%提升至第二阶段的15.3%。在基准测试中，Soofi S在所有完全开源模型中取得英语和德语综合最高分，超越OLMo 3 32B和Apertus 70B。在HumanEval上得分73.8%，MBPP得分70.2，德语版MBPP得分84.2。上下文窗口支持最高100万token，在4万token长度下，生成吞吐量约为同规模稠密模型的8倍。模型权重已开源。","category":"模型更新","source":"The Decoder：AI News（RSS）","pageTitle":"德国AI协会发布开源模型Soofi S，在英语和德语基准测试中领先 - Aioga AI资讯","description":"德国AI协会协调的研究联盟发布开源大语言模型Soofi S 30B-A3B。该模型总参数量316亿，每个token仅激活约32亿参数，采用Mamba-2与标准注意力层混合的MoE架构。模型完全在德国电信慕尼黑工业AI云上训练，训练数据中德语占比从第一阶段的7.2%提升至第二阶段的15.3%。在基准测试中，Soofi S在所有完全开源模型中取得英语和德语综合最","url":"https://www.aioga.com/news/cmrj6actv0651bilkm5pfz6ub/"},"en":{"title":"The German AI Association released the open-source model Soofi S, leading in English and German benchmark tests","summary":"The research consortium coordinated by the German AI Association has released the open-source large language model Soofi S 30B-A3B. The model has a total of 31.6 billion parameters, with each token activating only about 3.2 billion parameters, using a MoE architecture that combines Mamba-2 with a standard attention layer. The model is fully trained on Deutsche Telekom's Munich Industrial AI Cloud, with German accounting for training data increasing from 7.2% in the first phase to 15.3% in the second phase. In benchmarks, Soofi S achieved the highest scores in both English and German among all fully open-source models, surpassing OLMo 3 32B and Apertus 70B. It scored 73.8% on HumanEval, 70.2 on MBPP, and 84.2 on the German MBPP. The context window supports up to 1 million tokens, and at a length of 40,000 tokens, the generation throughput is about eight times that of a densely packed model of similar size. Model weights are open source.","category":"Models","source":"The Decoder：AI News（RSS）","pageTitle":"The German AI Association released the open-source model Soofi S, leading in English and German benchmark tests - Aioga AI News","description":"The research consortium coordinated by the German AI Association has released the open-source large language model Soofi S 30B-A3B. The model has a total of 31.6 billion parameters","url":"https://www.aioga.com/en/news/cmrj6actv0651bilkm5pfz6ub/","contentTranslated":true,"sourceHash":"019ccda520a73eae","translatedAt":"2026-07-19T15:57:50.995Z"},"ja":{"title":"ドイツAI協会はオープンソースモデルSoofi Sをリリースし、英語およびドイツ語のベンチマークテストでリードしています","summary":"ドイツAI協会が調整する研究コンソーシアムは、オープンソースの大規模言語モデルSoofi S 30B-A3Bをリリースしました。 このモデルは合計316億のパラメータを持ち、各トークンは約32億のパラメータしか有効化しません。これはMamba-2と標準的な注意層を組み合わせたMoEアーキテクチャを使用しています。 このモデルはドイツテレコムのミュンヘン工業AIクラウド上で完全に訓練されており、ドイツの学習データの占有率は第1フェーズの7.2%から第2フェーズの15.3%に増加しています。 ベンチマークでは、Soofi Sは全オープンソースモデルの中で英語とドイツ語の両方で最高得点を獲得し、OLMo 3 32BやApertus 70Bを上回りました。 HumanEvalで73.8%、MBPPで70.2、ドイツのMBPPで84.2のスコアを獲得しました。 コンテキストウィンドウは最大100万トークンをサポートし、4万トークンの長さでは、同じ規模の密集型モデルの約8倍の生成スループットとなります。 モデルの重みはオープンソースです。","category":"モデル更新","source":"The Decoder：AI News（RSS）","pageTitle":"ドイツAI協会はオープンソースモデルSoofi Sをリリースし、英語およびドイツ語のベンチマークテストでリードしています - Aioga AIニュース","description":"ドイツAI協会が調整する研究コンソーシアムは、オープンソースの大規模言語モデルSoofi S 30B-A3Bをリリースしました。 このモデルは合計316億のパラメータを持ち、各トークンは約32億のパラメータしか有効化しません。これはMamba-2と標準的な注意層を組み合わせたMoEアーキテクチャを使用しています。 このモデルはドイツテレコムのミュンヘン工業A","url":"https://www.aioga.com/ja/news/cmrj6actv0651bilkm5pfz6ub/","contentTranslated":true,"sourceHash":"019ccda520a73eae","translatedAt":"2026-07-19T15:57:51.458Z"},"ko":{"title":"독일 AI 협회는 오픈 소스 모델 Soofi S를 발표하며 영어와 독일어 벤치마크 테스트에서 선두를 달리고 있습니다","summary":"독일 AI 협회가 주관하는 연구 컨소시엄이 오픈 소스 대형 언어 모델 Soofi S 30B-A3B를 공개했습니다. 이 모델은 총 316억 개의 매개변수를 포함하며, 각 토큰은 약 32억 개의 매개변수만 활성화하며, Mamba-2와 표준 주의 계층을 결합한 MoE 아키텍처를 사용합니다. 이 모델은 도이체 텔레콤의 뮌헨 산업 AI 클라우드에서 완전히 학습되었으며, 독일 학습 데이터 비율은 1단계에서 7.2%에서 2단계에서는 15.3%로 증가했습니다. 벤치마크에서 Soofi S는 모든 완전 오픈 소스 모델 중 영어와 독일어 모두에서 가장 높은 점수를 기록하며 OLMo 3 32B와 Apertus 70B를 능가했습니다. HumanEval에서 73.8%, MBPP에서 70.2%, 독일 MBPP에서 84.2점을 받았습니다. 컨텍스트 윈도우는 최대 100만 개의 토큰을 지원하며, 4만 개의 토큰 길이에서 생성 처리량은 비슷한 크기의 밀집된 모델의 약 8배에 달합니다. 모델 가중치는 오픈 소스입니다.","category":"모델 업데이트","source":"The Decoder：AI News（RSS）","pageTitle":"독일 AI 협회는 오픈 소스 모델 Soofi S를 발표하며 영어와 독일어 벤치마크 테스트에서 선두를 달리고 있습니다 - Aioga AI 뉴스","description":"독일 AI 협회가 주관하는 연구 컨소시엄이 오픈 소스 대형 언어 모델 Soofi S 30B-A3B를 공개했습니다. 이 모델은 총 316억 개의 매개변수를 포함하며, 각 토큰은 약 32억 개의 매개변수만 활성화하며, Mamba-2와 표준 주의 계층을 결합한 MoE 아키텍처를 사용합니다. 이 모델은 도이체 텔레콤의 뮌헨 산업","url":"https://www.aioga.com/ko/news/cmrj6actv0651bilkm5pfz6ub/","contentTranslated":true,"sourceHash":"019ccda520a73eae","translatedAt":"2026-07-19T15:57:51.236Z"},"es":{"title":"La Asociación Alemana de IA lanzó el modelo de código abierto Soofi S, líder en pruebas de referencia en inglés y alemán","summary":"El consorcio de investigación coordinado por la Asociación Alemana de IA ha lanzado el modelo de lenguaje de código abierto Soofi S 30B-A3B. El modelo tiene un total de 31.600 millones de parámetros, con cada token activando solo unos 3.200 millones de parámetros, utilizando una arquitectura MoE que combina Mamba-2 con una capa de atención estándar. El modelo está completamente entrenado en la nube industrial de IA de Deutsche Telekom en Múnich, con la contabilidad alemana de los datos de entrenamiento que aumentó del 7,2% en la primera fase al 15,3% en la segunda. En los benchmarks, Soofi S obtuvo las puntuaciones más altas tanto en inglés como en alemán entre todos los modelos totalmente de código abierto, superando a OLMo 3 32B y Apertus 70B. Obtuvo un 73,8% en HumanEval, 70,2 en MBPP y 84,2 en MBPP alemán. La ventana de contexto soporta hasta 1 millón de tokens y, con una longitud de 40.000 tokens, el rendimiento de generación es aproximadamente ocho veces mayor que el de un modelo densamente empaquetado de tamaño similar. Los pesos de los modelos son de código abierto.","category":"Modelos","source":"The Decoder：AI News（RSS）","pageTitle":"La Asociación Alemana de IA lanzó el modelo de código abierto Soofi S, líder en pruebas de referencia en inglés y alemán - Aioga Noticias de IA","description":"El consorcio de investigación coordinado por la Asociación Alemana de IA ha lanzado el modelo de lenguaje de código abierto Soofi S 30B-A3B. El modelo tiene un total de 31.600 mill","url":"https://www.aioga.com/es/news/cmrj6actv0651bilkm5pfz6ub/","contentTranslated":true,"sourceHash":"019ccda520a73eae","translatedAt":"2026-07-19T15:57:51.264Z"},"fr":{"title":"L’Association allemande de l’IA a publié le modèle open source Soofi S, leader dans les tests de benchmark anglais et allemand","summary":"Le consortium de recherche, coordonné par l’Association allemande de l’IA, a publié le grand modèle de langage open source Soofi S 30B-A3B. Le modèle compte un total de 31,6 milliards de paramètres, chaque jeton n’activant qu’environ 3,2 milliards de paramètres, utilisant une architecture MoE qui combine Mamba-2 avec une couche d’attention standard. Le modèle est entièrement entraîné sur le cloud industriel d’IA de Munich de Deutsche Telekom, les données d’entraînement allemandes passant de 7,2 % lors de la première phase à 15,3 % lors de la seconde. Dans les benchmarks, Soofi S a obtenu les meilleurs scores en anglais et en allemand parmi tous les modèles entièrement open source, dépassant OLMo 3 32B et Apertus 70B. Il a obtenu 73,8 % sur HumanEval, 70,2 sur le MBPP et 84,2 sur le MBPP allemand. La fenêtre de contexte supporte jusqu’à 1 million de tokens, et avec une longueur de 40 000 tokens, le débit de génération est environ huit fois supérieur à celui d’un modèle densément compacté de taille similaire. Les poids des modèles sont open source.","category":"Modèles","source":"The Decoder：AI News（RSS）","pageTitle":"L’Association allemande de l’IA a publié le modèle open source Soofi S, leader dans les tests de benchmark anglais et allemand - Aioga Actualités IA","description":"Le consortium de recherche, coordonné par l’Association allemande de l’IA, a publié le grand modèle de langage open source Soofi S 30B-A3B. Le modèle compte un total de 31,6 millia","url":"https://www.aioga.com/fr/news/cmrj6actv0651bilkm5pfz6ub/","contentTranslated":true,"sourceHash":"019ccda520a73eae","translatedAt":"2026-07-19T15:57:51.354Z"},"de":{"title":"Der Deutsche KI-Verband veröffentlichte das Open-Source-Modell Soofi S, das bei englischen und deutschen Benchmark-Tests führend ist","summary":"Das vom Deutschen KI-Verband koordinierte Forschungskonsortium hat das Open-Source-Großsprachmodell Soofi S 30B-A3B veröffentlicht. Das Modell hat insgesamt 31,6 Milliarden Parameter, wobei jeder Token nur etwa 3,2 Milliarden Parameter aktiviert, wobei eine MoE-Architektur verwendet wird, die Mamba-2 mit einer Standard-Aufmerksamkeitsschicht kombiniert. Das Modell ist vollständig auf der Münchner Industrial AI Cloud der Deutschen Telekom trainiert, wobei die deutschen Trainingsdaten von 7,2 % in der ersten Phase auf 15,3 % in der zweiten Phase steigen. In Benchmarks erzielte Soofi S die höchsten Werte sowohl in Englisch als auch Deutsch unter allen vollständig Open-Source-Modellen und übertraf OLMo 3 32B und Apertus 70B. Sie erreichte 73,8 % auf HumanEval, 70,2 % auf MBPP und 84,2 % auf der deutschen MBPP. Das Kontextfenster unterstützt bis zu 1 Million Tokens, und bei einer Länge von 40.000 Token beträgt der Erzeugungsdurchsatz etwa achtmal so hoch wie bei einem dicht gepackten Modell ähnlicher Größe. Modellgewichte sind Open Source.","category":"模型更新","source":"The Decoder：AI News（RSS）","pageTitle":"Der Deutsche KI-Verband veröffentlichte das Open-Source-Modell Soofi S, das bei englischen und deutschen Benchmark-Tests führend ist - Aioga KI-News","description":"Das vom Deutschen KI-Verband koordinierte Forschungskonsortium hat das Open-Source-Großsprachmodell Soofi S 30B-A3B veröffentlicht. Das Modell hat insgesamt 31,6 Milliarden Paramet","url":"https://www.aioga.com/de/news/cmrj6actv0651bilkm5pfz6ub/","contentTranslated":true,"sourceHash":"019ccda520a73eae","translatedAt":"2026-07-19T15:57:55.359Z"},"pt-BR":{"title":"A Associação Alemã de IA lançou o modelo open-source Soofi S, líder em testes de benchmark em inglês e alemão","summary":"O consórcio de pesquisa coordenado pela Associação Alemã de IA lançou o modelo de linguagem de grande porte open-source Soofi S 30B-A3B. O modelo possui um total de 31,6 bilhões de parâmetros, com cada token ativando apenas cerca de 3,2 bilhões de parâmetros, usando uma arquitetura MoE que combina o Mamba-2 com uma camada padrão de atenção. O modelo é totalmente treinado na Nuvem Industrial de IA de Munique da Deutsche Telekom, com a contabilidade alemã dos dados de treinamento aumentando de 7,2% na primeira fase para 15,3% na segunda fase. Nos benchmarks, o Soofi S alcançou as maiores pontuações tanto em inglês quanto em alemão entre todos os modelos totalmente open-source, superando OLMo 3 32B e Apertus 70B. Obteve 73,8% no HumanEval, 70,2% no MBPP e 84,2% no MBPP alemão. A janela de contexto suporta até 1 milhão de tokens e, com um comprimento de 40.000 tokens, a taxa de geração é cerca de oito vezes maior que a de um modelo densamente compacto de tamanho semelhante. Os pesos dos modelos são open source.","category":"模型更新","source":"The Decoder：AI News（RSS）","pageTitle":"A Associação Alemã de IA lançou o modelo open-source Soofi S, líder em testes de benchmark em inglês e alemão - Aioga Notícias de IA","description":"O consórcio de pesquisa coordenado pela Associação Alemã de IA lançou o modelo de linguagem de grande porte open-source Soofi S 30B-A3B. O modelo possui um total de 31,6 bilhões de","url":"https://www.aioga.com/pt-BR/news/cmrj6actv0651bilkm5pfz6ub/","contentTranslated":true,"sourceHash":"019ccda520a73eae","translatedAt":"2026-07-19T15:57:55.819Z"},"ru":{"title":"Немецкая ассоциация искусственного интеллекта выпустила модель с открытым исходным кодом Soofi S, лидирующую в тестировании бенчмарков на английском и немецком языках","summary":"Исследовательский консорциум, координируемый Немецкой ассоциацией искусственного интеллекта, выпустил открытую большую языковую модель Soofi S 30B-A3B. Модель содержит в общей сложности 31,6 миллиарда параметров, при этом каждый токен активирует только около 3,2 миллиарда параметров, используя архитектуру MoE, объединяющую Mamba-2 со стандартным слоем внимания. Модель полностью обучена на Munich Industrial AI Cloud компании Deutsche Telekom, при этом немецкий учет обучающих данных увеличился с 7,2% на первом этапе до 15,3% на втором. В бенчмарках Soofi S показала наивысшие оценки как на английском, так и на немецком языках среди всех полностью открытых моделей, превзойдя OLMo 3 32B и Apertus 70B. Он набрал 73,8% на HumanEval, 70,2 на MBPP и 84,2 на немецком MBPP. Окно контекста поддерживает до 1 миллиона токенов, а при длине 40 000 токенов пропускная способность генерации примерно в восемь раз превышает плотно упакованную модель аналогичного размера. Веса моделей — это открытый исходный код.","category":"模型更新","source":"The Decoder：AI News（RSS）","pageTitle":"Немецкая ассоциация искусственного интеллекта выпустила модель с открытым исходным кодом Soofi S, лидирующую в тестировании бенчмарков на английском и немецком языках - Aioga Новости ИИ","description":"Исследовательский консорциум, координируемый Немецкой ассоциацией искусственного интеллекта, выпустил открытую большую языковую модель Soofi S 30B-A3B. Модель содержит в общей слож","url":"https://www.aioga.com/ru/news/cmrj6actv0651bilkm5pfz6ub/","contentTranslated":true,"sourceHash":"019ccda520a73eae","translatedAt":"2026-07-19T15:57:56.701Z"},"ar":{"title":"أصدرت جمعية الذكاء الاصطناعي الألمانية نموذج Soofi S مفتوح المصدر، الذي يتصدر اختبارات المعيار بالإنجليزية والألمانية","summary":"أصدر اتحاد البحث الذي تنسقته الجمعية الألمانية للذكاء الاصطناعي نموذج لغة كبيرة مفتوح المصدر Soofi S 30B-A3B. يحتوي النموذج على ما مجموعه 31.6 مليار معلم، حيث يفعل كل رمز حوالي 3.2 مليار معلم فقط، باستخدام بنية MoE التي تجمع بين مامبا-2 وطبقة انتباه قياسية. تم تدريب النموذج بالكامل على سحابة الذكاء الاصطناعي الصناعية التابعة لشركة دويتشه تيليكوم في ميونيخ، حيث ارتفعت بيانات التدريب في ألمانيا من 7.2٪ في المرحلة الأولى إلى 15.3٪ في المرحلة الثانية. في اختبارات المعيار، حققت Soofi S أعلى الدرجات في كل من اللغتين الإنجليزية والألمانية بين جميع النماذج مفتوحة المصدر بالكامل، متجاوزة OLMo 3 32B و Apertus 70B. حصلت على تقييم 73.8٪ في HumanEval، و70.2 على MBPP، و84.2 في MBPP الألماني. تدعم نافذة السياق ما يصل إلى مليون رمز، وبطول 40,000 رمز، فإن معدل التوليد حوالي ثمانية أضعاف معدل النموذج المكتظ بحجم مماثل. أوزان النماذج مفتوحة المصدر.","category":"模型更新","source":"The Decoder：AI News（RSS）","pageTitle":"أصدرت جمعية الذكاء الاصطناعي الألمانية نموذج Soofi S مفتوح المصدر، الذي يتصدر اختبارات المعيار بالإنجليزية والألمانية - Aioga أخبار الذكاء الاصطناعي","description":"أصدر اتحاد البحث الذي تنسقته الجمعية الألمانية للذكاء الاصطناعي نموذج لغة كبيرة مفتوح المصدر Soofi S 30B-A3B. يحتوي النموذج على ما مجموعه 31.6 مليار معلم، حيث يفعل كل رمز حوالي 3.2","url":"https://www.aioga.com/ar/news/cmrj6actv0651bilkm5pfz6ub/","contentTranslated":true,"sourceHash":"019ccda520a73eae","translatedAt":"2026-07-19T15:57:56.688Z"},"hi":{"title":"जर्मन एआई एसोसिएशन ने ओपन-सोर्स मॉडल सूफी एस जारी किया, जो अंग्रेजी और जर्मन बेंचमार्क परीक्षणों में अग्रणी है","summary":"जर्मन एआई एसोसिएशन द्वारा समन्वित अनुसंधान संघ ने ओपन-सोर्स लार्ज लैंग्वेज मॉडल सूफी एस 30बी-ए3बी जारी किया है। मॉडल में कुल 31.6 बिलियन पैरामीटर हैं, प्रत्येक टोकन केवल 3.2 बिलियन मापदंडों को सक्रिय करता है, एक MoE आर्किटेक्चर का उपयोग करता है जो माम्बा -2 को एक मानक ध्यान परत के साथ जोड़ता है। मॉडल को ड्यूश टेलीकॉम के म्यूनिख इंडस्ट्रियल एआई क्लाउड पर पूरी तरह से प्रशिक्षित किया गया है, जिसमें प्रशिक्षण डेटा के लिए जर्मन लेखांकन पहले चरण में 7.2% से बढ़कर दूसरे चरण में 15.3% हो गया है। बेंचमार्क में, सूफी एस ने सभी पूरी तरह से ओपन-सोर्स मॉडल के बीच अंग्रेजी और जर्मन दोनों में उच्चतम स्कोर हासिल किए, ओएलएमओ 3 32बी और एपर्टस 70बी को पीछे छोड़ दिया। इसने ह्यूमनइवल पर 73.8%, एमबीपीपी पर 70.2 और जर्मन एमबीपीपी पर 84.2 अंक प्राप्त किए। संदर्भ विंडो 1 मिलियन टोकन तक का समर्थन करती है, और 40,000 टोकन की लंबाई पर, पीढ़ी का थ्रूपुट समान आकार के घने पैक किए गए मॉडल से लगभग आठ गुना अधिक है। मॉडल वजन खुला स्रोत हैं।","category":"模型更新","source":"The Decoder：AI News（RSS）","pageTitle":"जर्मन एआई एसोसिएशन ने ओपन-सोर्स मॉडल सूफी एस जारी किया, जो अंग्रेजी और जर्मन बेंचमार्क परीक्षणों में अग्रणी है - Aioga AI समाचार","description":"जर्मन एआई एसोसिएशन द्वारा समन्वित अनुसंधान संघ ने ओपन-सोर्स लार्ज लैंग्वेज मॉडल सूफी एस 30बी-ए3बी जारी किया है। मॉडल में कुल 31.6 बिलियन पैरामीटर हैं, प्रत्येक टोकन केवल 3.2 बिलियन","url":"https://www.aioga.com/hi/news/cmrj6actv0651bilkm5pfz6ub/","contentTranslated":true,"sourceHash":"019ccda520a73eae","translatedAt":"2026-07-19T15:57:56.433Z"},"it":{"title":"L'Associazione Tedesca per l'IA ha rilasciato il modello open-source Soofi S, leader nei test di benchmark in inglese e tedesco","summary":"Il consorzio di ricerca coordinato dall'Associazione Tedesca per l'IA ha rilasciato il modello linguistico open source Soofi S 30B-A3B. Il modello ha un totale di 31,6 miliardi di parametri, con ogni token che attiva solo circa 3,2 miliardi di parametri, utilizzando un'architettura MoE che combina Mamba-2 con uno standard strato di attenzione. Il modello è completamente addestrato sul Munich Industrial AI Cloud di Deutsche Telekom, con la contabilizzazione tedesca dei dati di addestramento che è aumentata dal 7,2% nella prima fase al 15,3% nella seconda fase. Nei benchmark, Soofi S ha ottenuto i punteggi più alti sia in inglese che in tedesco tra tutti i modelli completamente open-source, superando OLMo 3 32B e Apertus 70B. Ha ottenuto il 73,8% su HumanEval, 70,2 su MBPP e 84,2% su MBPP tedesco. La finestra di contesto supporta fino a 1 milione di token, e con una lunghezza di 40.000 token, la capacità di generazione è circa otto volte quella di un modello densamente compatto di dimensioni simili. I pesi dei modelli sono open source.","category":"模型更新","source":"The Decoder：AI News（RSS）","pageTitle":"L'Associazione Tedesca per l'IA ha rilasciato il modello open-source Soofi S, leader nei test di benchmark in inglese e tedesco - Aioga Notizie IA","description":"Il consorzio di ricerca coordinato dall'Associazione Tedesca per l'IA ha rilasciato il modello linguistico open source Soofi S 30B-A3B. Il modello ha un totale di 31,6 miliardi di ","url":"https://www.aioga.com/it/news/cmrj6actv0651bilkm5pfz6ub/","contentTranslated":true,"sourceHash":"019ccda520a73eae","translatedAt":"2026-07-19T15:57:55.786Z"},"nl":{"title":"De Duitse AI-vereniging bracht het open-source model Soofi S uit, dat vooroploopt in Engelse en Duitse benchmarktests","summary":"Het onderzoeksconsortium, gecoördineerd door de Duitse AI-vereniging, heeft het open-source groottalenmodel Soofi S 30B-A3B uitgebracht. Het model heeft in totaal 31,6 miljard parameters, waarbij elke token slechts ongeveer 3,2 miljard parameters activeert, met een MoE-architectuur die Mamba-2 combineert met een standaard aandachtslaag. Het model is volledig getraind op Deutsche Telekom's Munich Industrial AI Cloud, waarbij de Duitse rekening voor trainingsdata steeg van 7,2% in de eerste fase tot 15,3% in de tweede fase. In benchmarks behaalde Soofi S de hoogste scores in zowel Engels als Duits van alle volledig open source modellen, waarmee ze OLMo 3 32B en Apertus 70B overtrof. Het scoorde 73,8% op HumanEval, 70,2 op MBPP en 84,2% op het Duitse MBPP. Het contextvenster ondersteunt tot 1 miljoen tokens, en met een lengte van 40.000 tokens is de generatiedoorvoer ongeveer acht keer zo groot als die van een dicht beladen model van vergelijkbare grootte. Modelgewichten zijn open source.","category":"模型更新","source":"The Decoder：AI News（RSS）","pageTitle":"De Duitse AI-vereniging bracht het open-source model Soofi S uit, dat vooroploopt in Engelse en Duitse benchmarktests - Aioga AI-nieuws","description":"Het onderzoeksconsortium, gecoördineerd door de Duitse AI-vereniging, heeft het open-source groottalenmodel Soofi S 30B-A3B uitgebracht. Het model heeft in totaal 31,6 miljard para","url":"https://www.aioga.com/nl/news/cmrj6actv0651bilkm5pfz6ub/","contentTranslated":true,"sourceHash":"019ccda520a73eae","translatedAt":"2026-07-19T15:57:55.790Z"},"tr":{"title":"Alman Yapay Zeka Derneği, İngilizce ve Almanca benchmark testlerinde lider olan açık kaynak modeli Soofi S'yi piyasaya sürdü","summary":"Alman Yapay Zeka Derneği tarafından koordine edilen araştırma konsorsiyumu, açık kaynaklı büyük dil modeli Soofi S 30B-A3B'yi yayımladı. Modelin toplamda 31,6 milyar parametresi vardır ve her token yalnızca yaklaşık 3,2 milyar parametreyi aktive eder; bu da Mamba-2'yi standart dikkat katmanıyla birleştiren bir MoE mimarisi kullanır. Model, Deutsche Telekom'un Münih Endüstriyel Yapay Zeka Bulutu üzerinde tamamen eğitilmiştir; alman eğitim verisi ilk aşamada %7,2'den ikinci aşamada %15,3'e yükselmiştir. Benchmarklarda, Soofi S tamamen açık kaynak modeller arasında hem İngilizce hem de Almanca en yüksek puanı elde ederek OLMo 3 32B ve Apertus 70B'yi geride bıraktı. HumanEval'de %73,8, MBPP'de 70,2 ve Alman MBPP'de 84,2 puan aldı. Bağlam penceresi 1 milyona kadar tokenı destekler ve 40.000 token uzunluğunda, üretim verimliliği benzer boyutta yoğun bir modelin yaklaşık sekiz katıdır. Model ağırlıkları açık kaynaklıdır.","category":"模型更新","source":"The Decoder：AI News（RSS）","pageTitle":"Alman Yapay Zeka Derneği, İngilizce ve Almanca benchmark testlerinde lider olan açık kaynak modeli Soofi S'yi piyasaya sürdü - Aioga AI Haberleri","description":"Alman Yapay Zeka Derneği tarafından koordine edilen araştırma konsorsiyumu, açık kaynaklı büyük dil modeli Soofi S 30B-A3B'yi yayımladı. Modelin toplamda 31,6 milyar parametresi va","url":"https://www.aioga.com/tr/news/cmrj6actv0651bilkm5pfz6ub/","contentTranslated":true,"sourceHash":"019ccda520a73eae","translatedAt":"2026-07-19T15:57:55.987Z"},"vi":{"title":"Hiệp hội AI Đức đã phát hành mô hình mã nguồn mở Soofi S, dẫn đầu trong các bài kiểm tra điểm chuẩn tiếng Anh và tiếng Đức","summary":"Hiệp hội nghiên cứu do Hiệp hội AI Đức điều phối đã phát hành mô hình ngôn ngữ lớn mã nguồn mở Soofi S 30B-A3B. Mô hình có tổng cộng 31,6 tỷ tham số, với mỗi token chỉ kích hoạt khoảng 3,2 tỷ tham số, sử dụng kiến trúc MoE kết hợp Mamba-2 với một lớp chú ý tiêu chuẩn. Mô hình được đào tạo đầy đủ trên Đám mây AI công nghiệp Munich của Deutsche Telekom, với dữ liệu đào tạo của Đức tăng từ 7,2% trong giai đoạn đầu lên 15,3% trong giai đoạn hai. Về điểm chuẩn, Soofi S đạt được điểm số cao nhất bằng cả tiếng Anh và tiếng Đức trong số tất cả các mô hình mã nguồn mở hoàn toàn, vượt qua OLMo 3 32B và Apertus 70B. Nó đạt 73,8% trên HumanEval, 70,2 trên MBPP và 84,2 trên MBPP của Đức. Cửa sổ ngữ cảnh hỗ trợ tối đa 1 triệu token và với độ dài 40.000 token, thông lượng tạo gấp khoảng tám lần so với một mô hình dày đặc có kích thước tương tự. Trọng lượng mô hình là mã nguồn mở.","category":"模型更新","source":"The Decoder：AI News（RSS）","pageTitle":"Hiệp hội AI Đức đã phát hành mô hình mã nguồn mở Soofi S, dẫn đầu trong các bài kiểm tra điểm chuẩn tiếng Anh và tiếng Đức - Tin tức AI Aioga","description":"Hiệp hội nghiên cứu do Hiệp hội AI Đức điều phối đã phát hành mô hình ngôn ngữ lớn mã nguồn mở Soofi S 30B-A3B. Mô hình có tổng cộng 31,6 tỷ tham số, với mỗi token chỉ kích hoạt kh","url":"https://www.aioga.com/vi/news/cmrj6actv0651bilkm5pfz6ub/","contentTranslated":true,"sourceHash":"019ccda520a73eae","translatedAt":"2026-07-19T15:57:56.408Z"},"id":{"title":"Asosiasi AI Jerman merilis model sumber terbuka Soofi S, yang memimpin dalam pengujian benchmark bahasa Inggris dan Jerman","summary":"Konsorsium penelitian yang dikoordinasikan oleh Asosiasi AI Jerman telah merilis model bahasa besar sumber terbuka Soofi S 30B-A3B. Model ini memiliki total 31,6 miliar parameter, dengan setiap token hanya mengaktifkan sekitar 3,2 miliar parameter, menggunakan arsitektur MoE yang menggabungkan Mamba-2 dengan lapisan perhatian standar. Model ini sepenuhnya dilatih di Munich Industrial AI Cloud Deutsche Telekom, dengan Jerman memperhitungkan data pelatihan meningkat dari 7,2% pada fase pertama menjadi 15,3% pada fase kedua. Dalam tolok ukur, Soofi S mencapai skor tertinggi dalam bahasa Inggris dan Jerman di antara semua model sumber terbuka sepenuhnya, melampaui OLMo 3 32B dan Apertus 70B. Itu mencetak 73,8% di HumanEval, 70,2 di MBPP, dan 84,2 di MBPP Jerman. Jendela konteks mendukung hingga 1 juta token, dan dengan panjang 40.000 token, throughput generasi sekitar delapan kali lipat dari model padat dengan ukuran yang sama. Bobot model bersifat open source.","category":"模型更新","source":"The Decoder：AI News（RSS）","pageTitle":"Asosiasi AI Jerman merilis model sumber terbuka Soofi S, yang memimpin dalam pengujian benchmark bahasa Inggris dan Jerman - Berita AI Aioga","description":"Konsorsium penelitian yang dikoordinasikan oleh Asosiasi AI Jerman telah merilis model bahasa besar sumber terbuka Soofi S 30B-A3B. Model ini memiliki total 31,6 miliar parameter, ","url":"https://www.aioga.com/id/news/cmrj6actv0651bilkm5pfz6ub/","contentTranslated":true,"sourceHash":"019ccda520a73eae","translatedAt":"2026-07-19T15:57:55.444Z"},"th":{"title":"สมาคม AI ของเยอรมันเปิดตัวโมเดลโอเพ่นซอร์ส Soofi S ซึ่งเป็นผู้นําในการทดสอบเกณฑ์มาตรฐานภาษาอังกฤษและเยอรมัน","summary":"สมาคมวิจัยที่ประสานงานโดยสมาคม AI ของเยอรมันได้เปิดตัวโมเดลภาษาขนาดใหญ่แบบโอเพ่นซอร์ส Soofi S 30B-A3B โมเดลนี้มีพารามิเตอร์ทั้งหมด 31.6 พันล้านตัว โดยแต่ละโทเค็นเปิดใช้งานพารามิเตอร์เพียงประมาณ 3.2 พันล้านตัว โดยใช้สถาปัตยกรรม MoE ที่รวม Mamba-2 เข้ากับเลเยอร์ความสนใจมาตรฐาน โมเดลนี้ได้รับการฝึกอบรมอย่างเต็มที่บน Munich Industrial AI Cloud ของ Deutsche Telekom โดยข้อมูลการฝึกอบรมของเยอรมันเพิ่มขึ้นจาก 7.2% ในระยะแรกเป็น 15.3% ในระยะที่สอง ในเกณฑ์มาตรฐาน Soofi S ได้รับคะแนนสูงสุดทั้งในภาษาอังกฤษและเยอรมันในบรรดาโมเดลโอเพ่นซอร์สเต็มรูปแบบ แซงหน้า OLMo 3 32B และ Apertus 70B ได้คะแนน 73.8% ใน HumanEval, 70.2 ใน MBPP และ 84.2 ใน MBPP ของเยอรมัน หน้าต่างบริบทรองรับโทเค็นได้มากถึง 1 ล้านโทเค็น และด้วยความยาว 40,000 โทเค็น ปริมาณงานของการสร้างจะมากกว่าโมเดลที่มีขนาดใกล้เคียงกันประมาณแปดเท่า น้ําหนักแบบจําลองเป็นโอเพ่นซอร์ส","category":"模型更新","source":"The Decoder：AI News（RSS）","pageTitle":"สมาคม AI ของเยอรมันเปิดตัวโมเดลโอเพ่นซอร์ส Soofi S ซึ่งเป็นผู้นําในการทดสอบเกณฑ์มาตรฐานภาษาอังกฤษและเยอรมัน - ข่าว AI Aioga","description":"สมาคมวิจัยที่ประสานงานโดยสมาคม AI ของเยอรมันได้เปิดตัวโมเดลภาษาขนาดใหญ่แบบโอเพ่นซอร์ส Soofi S 30B-A3B โมเดลนี้มีพารามิเตอร์ทั้งหมด 31.6 พันล้านตัว โดยแต่ละโทเค็นเปิดใช้งานพารามิเตอ","url":"https://www.aioga.com/th/news/cmrj6actv0651bilkm5pfz6ub/","contentTranslated":true,"sourceHash":"019ccda520a73eae","translatedAt":"2026-07-19T15:57:55.962Z"},"pl":{"title":"Niemieckie Stowarzyszenie AI opublikowało model open source Soofi S, który wyróżnia się w testach testów w języku angielskim i niemieckim","summary":"Konsorcjum badawcze koordynowane przez Niemieckie Stowarzyszenie AI opublikowało otwarty model językowy Soofi S 30B-A3B. Model ma łącznie 31,6 miliarda parametrów, z których każdy token aktywuje tylko około 3,2 miliarda parametrów, wykorzystując architekturę MoE łączącą Mamba-2 ze standardową warstwą uwagi. Model jest w pełni trenowany na Deutsche Telekom Munich Industrial AI Cloud, gdzie niemiecki odpowiada za wzrost danych treningowych z 7,2% w pierwszej fazie do 15,3% w drugiej. W testach benchmarkowych Soofi S uzyskał najwyższe wyniki zarówno w języku angielskim, jak i niemieckim spośród wszystkich w pełni otwartych modeli, wyprzedzając OLMo 3 32B i Apertus 70B. Uzyskał 73,8% w HumanEval, 70,2% w MBPP oraz 84,2% w niemieckim MBPP. Okno kontekstowe obsługuje do 1 miliona tokenów, a przy długości 40 000 tokenów przepustowość generowania jest około osiem razy większa niż w gęsto upakowanym modelu o podobnej wielkości. Wagi modeli są open source.","category":"模型更新","source":"The Decoder：AI News（RSS）","pageTitle":"Niemieckie Stowarzyszenie AI opublikowało model open source Soofi S, który wyróżnia się w testach testów w języku angielskim i niemieckim - Aioga Wiadomości AI","description":"Konsorcjum badawcze koordynowane przez Niemieckie Stowarzyszenie AI opublikowało otwarty model językowy Soofi S 30B-A3B. Model ma łącznie 31,6 miliarda parametrów, z których każdy ","url":"https://www.aioga.com/pl/news/cmrj6actv0651bilkm5pfz6ub/","contentTranslated":true,"sourceHash":"019ccda520a73eae","translatedAt":"2026-07-19T15:57:57.047Z"}}}}