File size: 13,308 Bytes
7b117b8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
778f7ff
7b117b8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
55a438a
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
<!DOCTYPE html>
<html lang="en">
<head>
  <meta charset="utf-8">
  <meta name="description" content="Soofi S 30B-A3B: A sovereign, open-source foundation model for German and English.">
  <meta name="viewport" content="width=device-width, initial-scale=1">
  <title>Soofi S: A Sovereign, Open-Source Foundation Model for German and English</title>
  <link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" rel="stylesheet">
  <link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/bulma@0.9.4/css/bulma.min.css">
  <link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/5.15.4/css/all.min.css">
  <link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
  <style>
    body { font-family: 'Noto Sans', sans-serif; }
    /* soofipurple from the report: RGB(231,206,246) */
    .hero.is-primary { background-color: #e7cef6; }
    .hero.is-primary .title, .hero.is-primary .subtitle { color: #363636; }
    .publication-authors { line-height: 1.6; }
    .author-group { margin-top: 0.6em; }
    .stat-box { border: 1px solid #e0e0e0; border-radius: 8px; padding: 1em; height: 100%; }
    .stat-box .heading { color: #7a7a7a; }
    .link-block { margin: 0.15em; display: inline-block; }
    figure img { max-width: 100%; height: auto; }
  </style>
</head>
<body>

<section class="hero is-primary">
  <div class="hero-body">
    <div class="container has-text-centered">
      <!-- Optional: replace the emoji with the project wordmark, e.g. <img src="logo.png" style="max-width:340px"> -->
      <img src="logo.png" alt="Soofi S" style="max-width: 340px; height: auto;">
      <p class="subtitle is-5">A sovereign, open-source Mixture-of-Experts hybrid Mamba–Transformer foundation model for German and English</p>
    </div>
  </div>
</section>

<section class="hero">
  <div class="hero-body">
    <div class="container is-max-desktop">
      <div class="columns is-centered">
        <div class="column has-text-centered">
          <h1 class="title is-2 publication-title">A Sovereign, Open-Source Foundation Model for German and English</h1>
          <p class="subtitle is-5">Soofi S Pretraining Report v1.0</p>

          <div class="is-size-4 publication-authors">
            <p class="author-group"><strong>The Soofi-Team</strong></p>
          </div>

          <div class="is-size-6 publication-authors" style="margin-top: 1em;">
            <span class="author-block">KI Bundesverband,</span>
            <span class="author-block">DFKI,</span>
            <span class="author-block">Fraunhofer IAIS,</span>
            <span class="author-block">Fraunhofer IIS,</span>
            <span class="author-block">Technische Universität Darmstadt,</span>
            <span class="author-block">Universität Würzburg,</span>
            <span class="author-block">Berliner Hochschule für Technik,</span>
            <span class="author-block">L3S Research Center,</span>
            <span class="author-block">Lamarr,</span>
            <span class="author-block">ellamind,</span>
            <span class="author-block">hessian.AI,</span>
            <span class="author-block">Merantix Momentum</span>
          </div>

          <p class="is-size-6" style="margin-top: 1em;">
            Consortium coordinated by the KI Bundesverband. Funded by the German Federal Ministry for
            Economic Affairs and Energy (BMWE) in the context of IPCEI-CIS and 8ra.
          </p>

          <div class="column has-text-centered" style="margin-top: 0.5em;">
            <span class="link-block">
              <a href="https://arxiv.org/abs/2607.09424" target="_blank" class="external-link button is-normal is-rounded is-dark">
                <span class="icon"><i class="ai ai-arxiv"></i></span>
                <span>arXiv</span>
              </a>
            </span>
            <span class="link-block">
              <a href="https://huggingface.co/Soofi-Project" target="_blank" class="external-link button is-normal is-rounded is-dark">
                <span class="icon"><img src="https://huggingface.co/front/assets/huggingface_logo.svg" alt="Hugging Face" style="height: 1.0em; vertical-align: middle;"></span>
                <span>Models &amp; Checkpoints</span>
              </a>
            </span>
            <span class="link-block">
              <a href="https://github.com/soofi-project/Soofi-Pretraining" target="_blank" class="external-link button is-normal is-rounded is-dark">
                <span class="icon"><i class="fab fa-github"></i></span>
                <span>Training &amp; Data Code</span>
              </a>
            </span>
            <span class="link-block">
              <a href="https://github.com/ellamind/base-eval" target="_blank" class="external-link button is-normal is-rounded is-dark">
                <span class="icon"><i class="fab fa-github"></i></span>
                <span>Eval: base-eval</span>
              </a>
            </span>
            <span class="link-block">
              <a href="https://github.com/ellamind/eval-hive" target="_blank" class="external-link button is-normal is-rounded is-dark">
                <span class="icon"><i class="fab fa-github"></i></span>
                <span>Eval: eval-hive</span>
              </a>
            </span>
            <span class="link-block">
              <a href="https://api.wandb.ai/links/soofi-exchange/j11vi7rg" target="_blank" class="external-link button is-normal is-rounded is-dark">
                <span class="icon"><i class="fas fa-chart-line"></i></span>
                <span>Training Logs (W&amp;B)</span>
              </a>
            </span>
          </div>

          <div class="notification is-warning is-light" style="margin-top: 1em;">
            <strong>Note:</strong> During the current beta phase, the model repositories are gated —
            you need to accept the access conditions on Hugging Face before downloading.
            Once the beta phase ends, the models will be freely available without access request.
          </div>
        </div>
      </div>
    </div>
  </div>
</section>

<section class="section">
  <div class="container content is-max-desktop">
    <p>
      <strong>Soofi S 30B-A3B</strong> is a sovereign, open-source Mixture-of-Experts (MoE) hybrid Mamba–Transformer
      foundation model for German and English. Its hybrid design activates only 3B of 30B parameters per token and
      keeps the inference cache near-constant as context grows, giving it a decisive throughput advantage over dense
      models for long-context, high-concurrency deployment.
    </p>
    <p>
      Pretrained on roughly 27 trillion tokens with deliberately up-weighted German, Soofi S matches dense 14–27B
      models on aggregate English and German benchmarks, achieves the best code aggregates in both languages among
      17 open base models, and outperforms every European sovereign baseline in our comparison — including ones far
      larger in active parameters. Among fully open models, it obtains the highest English and German evaluation
      scores, ahead of Olmo 3 32B and Apertus 70B. Soofi S was built end-to-end on the German Industrial AI Cloud,
      a sovereign HPC-scale AI infrastructure operated by Deutsche Telekom in Munich.
    </p>
  </div>
</section>

<section class="section pt-0">
  <div class="container is-max-desktop">
    <h2 class="title is-3">📌 At a Glance</h2>
    <div class="columns is-multiline">
      <div class="column is-3"><div class="stat-box has-text-centered"><p class="heading">Parameters</p><p class="title is-4">31.6B total<br>3.2B active</p></div></div>
      <div class="column is-3"><div class="stat-box has-text-centered"><p class="heading">Architecture</p><p class="title is-4">52 layers<br>23 Mamba-2 · 23 MoE · 6 GQA</p></div></div>
      <div class="column is-3"><div class="stat-box has-text-centered"><p class="heading">Pretraining</p><p class="title is-4">~26.68T tokens<br>DE up to 15.3%</p></div></div>
      <div class="column is-3"><div class="stat-box has-text-centered"><p class="heading">Context</p><p class="title is-4">up to 1M<br>tokens</p></div></div>
    </div>
  </div>
</section>

<section class="section">
  <div class="container content is-max-desktop">
    <h2 class="title is-3">🧩 Highlights</h2>
    <ol>
      <li><strong>🏆 German–English champion:</strong> best English and German code aggregates among all 17 measured
        open base models; strongest fully open model on the English and German aggregates; matches or outperforms
        every European sovereign baseline on every German benchmark in our suite — at a fraction of the
        active-parameter cost of dense 14–27B models.</li>
      <li><strong>📋 Full data transparency:</strong> complete per-source, per-language token accounting for all
        three training phases (including sources we evaluated and <em>excluded</em>), with reproducible corpus
        construction scripts. ~99% of the mixture can be independently reconstructed.</li>
      <li><strong>🔁 Reproducible recipe:</strong> full Warmup–Stable–Decay learning-rate schedule, optimizer, all
        hyperparameters, per-phase token budgets, and phase boundaries — a third party can rebuild the run.</li>
      <li><strong>⚡ Long-context serving efficiency:</strong> only 6 of 52 layers keep a KV cache, so the
        per-sequence cache stays near-constant with context length. Measured aggregate decode TPS/GPU is 8–9× that
        of dense 14–24B models at 40K context (batch 32) and stays essentially flat from 4K to 256K.</li>
      <li><strong>🇩🇪 Sovereign end to end:</strong> trained from 24 March to 13 May 2026 on up to 512 NVIDIA B200
        GPUs of the German Industrial AI Cloud (Deutsche Telekom, Munich), under European operational and
        data-protection requirements.</li>
    </ol>
  </div>
</section>

<section class="section">
  <div class="container content is-max-desktop">
    <h2 class="title is-3">📊 Results</h2>
    <ul>
      <li><strong>Aggregates:</strong> English 70.1 / German 79.1 — highest among fully open models (Olmo 3 32B:
        67.3 / 69.2; Apertus 70B: 62.4 / 72.8), on par with dense open-weight 14–27B models.</li>
      <li><strong>Code:</strong> HumanEval 73.8, MBPP 70.2, MBPP-DE 84.2 — best in both comparison sets.</li>
      <li><strong>Mathematics:</strong> GSM8K 86.1, GSM8K-Platinum-DE 87.1, Minerva-500 79.4.</li>
      <li><strong>German:</strong> first on every German benchmark against European open-source baselines
        (GLP-DE 88.8, INCLUDE-DE 61.2, ARC-Challenge-DE 92.3).</li>
      <li><strong>Serving:</strong> 4.82k aggregate decode TPS/GPU at 40K context (batch 32, single B200, vLLM) —
        9.2× Ministral 3 14B — with near-flat decode throughput from 4K to 256K.</li>
    </ul>
  </div>
</section>

<section class="section">
  <div class="container content is-max-desktop">
    <h2 class="title is-3">📁 Available Artifacts</h2>
    <ul>
      <li><a href="https://huggingface.co/Soofi-Project" target="_blank">🤗 Soofi S 30B-A3B Base weights and selected intermediate checkpoints</a>
        <em>(currently gated during the beta phase; freely available once the beta ends)</em></li>
      <li><a href="https://github.com/soofi-project/Soofi-Pretraining" target="_blank">🛠️ Training code and reproducible data-construction scripts</a></li>
      <li><a href="https://github.com/ellamind/base-eval" target="_blank">🧪 Evaluation code (base-eval)</a> and <a href="https://github.com/ellamind/eval-hive" target="_blank">eval-hive</a>, incl. parallel English/German suites</li>
      <li><a href="https://api.wandb.ai/links/soofi-exchange/j11vi7rg" target="_blank">📈 Weights &amp; Biases dashboard of the full ~27T-token run</a></li>
      <li>📋 Exact per-source token accounting for all three phases; commercially licensed sources (Genios) documented via aggregate statistics</li>
    </ul>
    <p>All artifacts are released under permissive licenses for transparent audit and extension.</p>
  </div>
</section>

<section class="section">
  <div class="container content is-max-desktop">
    <h2 class="title is-3">📜 Citation</h2>
    <p>If you use Soofi S or its artifacts, please cite the report
      (<a href="https://arxiv.org/abs/2607.09424" target="_blank">arXiv:2607.09424</a>):</p>
    <pre><code>@misc{soofi2026soofis,
  title         = {A Sovereign, Open-Source Foundation Model for German and English: Soofi S Pretraining Report v1.0},
  author        = {{The Soofi-Team}},
  year          = {2026},
  eprint        = {2607.09424},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2607.09424}
}</code></pre>
  </div>
</section>

<footer class="footer">
  <div class="container content has-text-centered is-size-7">
    <p>
      Funded by the German Federal Ministry for Economic Affairs and Energy (BMWE) in the context of IPCEI-CIS and 8ra
      through “Soofi: Souveräne KI für Europa” (grant no. 13IPC040A-J).
      Website template adapted from the <a href="https://nerfies.github.io" target="_blank">Nerfies</a>/JQL project pages (Bulma).
    </p>
  </div>
</footer>

</body>
</html>