Skip to content
#small language model Open access

Deployment Matrix for Critical Infrastructure and Defence: Kolibri-1, Qwen3.6 and Mistral Small 4 on a Deterministic German Rule-Application Benchmark

Oct 2026 · Zenodo (CERN European Organization for Nuclear Research) · 2 references
Topic Modeling

Abstract

English Three language models (Aleph Alpha Kolibri-1, Qwen3.6-35B-A3B, Mistral Small 4) were compared on one H200 with vLLM 0.29.0 on a synthetic German rule-application benchmark. Each question comes with its source passages plus three distractors, so the test measures reading and rule application, not knowledge stored in the weights. Scoring is deterministic (Stage A); the semantic Stage B was not run. Main findings. With thinking enabled, Qwen3.6 detects 86% of conflicting regulations, Mistral Small 4 46% and Kolibri-1 39%. False statements on conflict and gap questions: Qwen 6.9%, Kolibri 12.5%, Mistral 23.7%. Without thinking, all three fail on conflicts (0–5% correct). Qwen needs the longest wait (26–53 s to first answer) and the most energy; Kolibri-1 is the most economical (143–197 J per correct answer with thinking). The pre-registered hypotheses H3 and H4 were not supported. There is no overall winner. Limits. Synthetic corpus, no real service regulations; Stage A only; unequal active parameters; Kolibri ran with an FP8 KV cache; origin, licence and NIS2/GDPR conformity of the models were not measured. This is decision support on test data, not an approval. Related work: Kolibri-1 Praxistest, Kolibri-1 Audit Report. Raw data, pre-registrations and SHA-256 checksums are included; the test and scoring scripts are withheld and available on request to authorised parties (see section "Evidence" in the report). Deutsch Drei Sprachmodelle (Aleph Alpha Kolibri-1, Qwen3.6-35B-A3B, Mistral Small 4) wurden auf einer H200 mit vLLM 0.29.0 an einem synthetischen deutschen Regelanwendungstest verglichen. Jede Frage enthält ihre Fundstellen plus drei Ablenker; gemessen werden also Lesen und Regelanwendung, nicht Wissen aus den Gewichten. Die Bewertung ist deterministisch (Stufe A); die semantische Stufe B wurde nicht ausgeführt. Hauptbefunde. Mit Nachdenken erkennt Qwen3.6 86 % der Vorschriftenwidersprüche, Mistral Small 4 46 % und Kolibri-1 39 %. Falsche Aussagen bei Widerspruchs- und Lückenfragen: Qwen 6,9 %, Kolibri 12,5 %, Mistral 23,7 %. Ohne Nachdenken scheitern alle drei bei Widersprüchen (0–5 % richtig). Qwen braucht die längste Wartezeit (26–53 s bis zur ersten Antwort) und die meiste Energie; Kolibri-1 ist am sparsamsten (143–197 J je richtiger Antwort mit Nachdenken). Die vorab registrierten Hypothesen H3 und H4 wurden nicht gestützt. Einen Gesamtsieger gibt es nicht. Grenzen. Synthetischer Korpus, keine echten Dienstvorschriften; nur Stufe A; ungleiche Aktivparameter; Kolibri lief mit FP8-KV-Cache; Herkunft, Lizenz und NIS2/DSGVO-Konformität der Modelle wurden nicht gemessen. Es ist eine Entscheidungshilfe auf Testdaten, keine Freigabe. Verwandte Arbeiten: Kolibri-1 Praxistest, Kolibri-1 Audit Report. Rohdaten, Vorab-Registrierungen und SHA-256-Prüfsummen liegen bei; die Test- und Bewertungsskripte werden zurückgehalten und auf Anfrage an autorisierte Personen herausgegeben (siehe Abschnitt „Nachweis“ im Bericht). الملخص (بالعربية) قارنت هذه الدراسة بين ثلاثة نماذج لغوية (Kolibri-1 من Aleph Alpha وQwen3.6-35B-A3B وMistral Small 4) على وحدة معالجة رسومية واحدة من طراز H200 باستخدام vLLM 0.29.0، وذلك في اختبار اصطناعي لتطبيق القواعد باللغة الألمانية. يتضمن كل سؤال المقاطع المصدرية ذات الصلة مع ثلاثة مشتتات، لذا يقيس الاختبار القراءة وتطبيق القواعد، لا المعرفة المخزّنة في الأوزان. التقييم حتمي (المرحلة أ)، ولم تُنفَّذ المرحلة الدلالية (المرحلة ب). عند تفعيل التفكير اكتشف Qwen3.6 نسبة 86% من التعارضات بين اللوائح، مقابل 46% لـ Mistral Small 4 و39% لـ Kolibri-1. وكانت نسبة العبارات الخاطئة في أسئلة التعارض والنقص 6.9% لـ Qwen و12.5% لـ Kolibri و23.7% لـ Mistral. وعند إيقاف التفكير فشلت النماذج الثلاثة جميعها في حالات التعارض (0 إلى 5% إجابات صحيحة). يحتاج Qwen إلى أطول زمن انتظار وأكبر قدر من الطاقة، بينما يتميّز Kolibri-1 بأقل استهلاك للطاقة. لم تُؤيَّد الفرضيتان H3 وH4 المسجّلتان مسبقاً، ولا يوجد فائز واحد شامل. حدود الدراسة: مدوّنة اصطناعية دون لوائح خدمة حقيقية، والمرحلة أ فقط، وتفاوت في عدد المعاملات النشطة، وتشغيل Kolibri بذاكرة KV بصيغة FP8. لم تُقَس جهة المنشأ ولا الترخيص ولا الامتثال لـ NIS2 وGDPR. والدراسة أداة مساعدة للقرار على بيانات اختبار، وليست موافقة أو اعتماداً.

View source

Similar papers

#small language model Dataset Open access Oct 2026

Socratic guiding questions in synthetic arithmetic data: matched LoRA runs (revision v2)

Supporting data, adapters, predictions and code for the article *Low-Cost LoRA Fine-Tuning of Small Language Models for Multi-Step Arithmetic Reasoning* by Jake O'Grady, Asena Isik Gürhan, Chee Fong Ting and Effirul Ramlan (University of Galway). We generated 20,000 GSM8K-derived arithmetic problems with step-by-step s...

O'Grady, Jake, Gürhan, Asena Isik, Chee, Fong Ting et al. · 465 citations
#computer vision Open access Jun 2016

Software Development in Startup Companies: The Greenfield Startup Model

The results are packaged in the Greenfield Startup Model (GSM), which explains the priority of startups to release the product as quickly as possible, and the need to shorten time-to-market, by speeding up the development through low-precision engineering activities.

Carmine Giardino, Nicolò Paternoster, M. Unterkalmsteiner et al. · 178 citations · ⚡14
#computer vision Open access Oct 2016

Software Startups - A Research Agenda

Software startup companies develop innovative, software-intensive products within limited timeframes and with few resources, searching for sustainable and scalable business models.

M. Unterkalmsteiner, P. Abrahamsson, Xiaofeng Wang et al. · 157 citations · ⚡17
#machine learning Review Open access Oct 2016

“Failures” to be celebrated: an analysis of major pivots of software startups

This study conducts a case survey study based on the secondary data of the major pivots happened in 49 software startups, and demonstrates that customer need pivot is the most common among all pivot types.

Sohaib Shahid Bajwa, Xiaofeng Wang, Anh Nguyen-Duc et al. · 127 citations · ⚡15
#computer vision Review Open access May 2015

A survey study on major technical barriers affecting the decision to adopt cloud services

The comparison of adopter and non-adopter sample reveals three potential adoption inhibitor, security, data privacy, and portability, which underlines the importance of the technical and security perspectives for research investigating the adoption of technology.

Nattakarn Phaphoom, Xiaofeng Wang, S. Samuel et al. · 111 citations · ⚡8
#computer vision Open access Feb 2018

Lean Internal Startups for Software Product Innovation in Large Companies: Enablers and Inhibitors

This study investigates how Lean internal startup facilitates software product innovation in large companies and identifies its enablers and inhibitors, and shows the potential of the method-in-action framework to investigate the Lean startup approach in non-startup context.

Henry Edison, Nina M. Smørsgård, Xiaofeng Wang et al. · 78 citations · ⚡6

Related blog posts

MIT News · Artificial Intelligence Oct 23, 2026

Q&A: How AI came to model the “messy” human mind

Assistant Professor Benjamin Lindquist studies how engineers deliberately injected randomness into early AI, giving rise to deep-learning models that sometimes hallucinate.

Microsoft Research Blog Oct 22, 2026

Flint-interactive: A language for creating interactive data visualizations

Flint-interactive gives AI agents a language for creating interactive visualizations. Click a chart, filter data, highlight related content, and explore beyond the view the author originally chose. The post Flint-interactive: A language for creating interactive data visualizations appeared first on Microsoft Research.

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.