Aleph Alpha's Kolibri gives European institutions something more concrete than a sovereignty pledge: downloadable weights they can retain, adapt and run inside infrastructure they choose. The German-English model became available October 3, with a further company announcement October 5. Its strongest proposition is control over deployment and German-language work, not global leadership on every benchmark.
Germany's government did not release Kolibri or select it as a national public-sector model. A private German company released a contender for administration, industry, defense, aerospace and other regulated work. Adoption still depends on evaluation, integration and procurement. Existing public-sector relationships are not evidence that this particular release already handles consequential government decisions.
Sparse computation does not mean a small model
Kolibri has 78.1 billion total parameters and activates about 3.46 billion per token. Its mixture-of-experts architecture selects six of 384 routed experts alongside one shared expert in each layer. Only ten of 50 layers use full attention; the others use a 512-token sliding window. These choices reduce computation, but the complete weights still need to remain accessible.
Apache 2.0 permits commercial use, modification and redistribution. BF16 and lower-precision FP8 releases are available; the latter's listed footprint is about 78 GB. Open weights provide portability beyond an API, not complete reproduction of every training decision. Full reconstruction still needs data, code, methods, expertise and substantial computing resources.
The model supports structured output, native tool calls and selectable reasoning: off, low, medium or high. An OpenAI-compatible serving interface lowers integration friction. Its model card positions it in advisory systems where people review output before action, not as an autonomous decision-maker for consequential cases.
The advertised maximum context is 1,048,576 tokens, while the native long-context training length is 262,144. Earlier stages used 16,384 and 65,536 tokens. Aleph Alpha recommends staying at or below the native length for complex tasks and efficient operation; the million-token setting was validated through extrapolation, not equivalent native training.
German specialization, with a numerical discrepancy
Aleph Alpha argues that German administrative writing requires more than fluent answers from an English-first model. Legal references, long compounds, institutional roles and procedural assumptions can be distorted by generic filtering or translation. It retuned German web filters and included parliamentary, legal and public-domain material.
The source accounts differ on the German share: two describe 21.3% of main pre-training, roughly 4.3 trillion token presentations within 20 trillion; another gives 23.9%, alongside 62.5% English and 13.6% code. One other account rounds English/code to 62%/14%. These percentages cannot be presented as one reconciled denominator. They are a documentation discrepancy to resolve, not an excuse to invent a corrected value.
The data account says filtering and deduplication initially produced 390 billion usable German tokens, short of an approximately 4 trillion-token target. A dedicated Common Crawl pipeline gathered 1.3 trillion unique tokens, and rephrasing German documents produced about 1 trillion more in formats such as questions and answers or encyclopedia passages. These are corpus construction and repeated presentation figures, not a license to add all counts as unique data.
A bilingual 128,000-token vocabulary uses the UniBPE approach for German morphology. Aleph Alpha reports 4.7 bytes per token in German and 4.2 in English, with more meaningful splits of administrative compounds. Better compression may save context and computation, but it cannot establish legal accuracy.
Approximately 800,000 German supervised reasoning examples used German sentence beginnings to reduce drift into English. A readable German explanation can help staff review an answer; generated reasoning is not necessarily a faithful trace of internal computation. Source citations, versioned prompts and access records remain necessary. The model's German and English knowledge cutoff is June 18, 2026, so newer facts need retrieval or tools.
The training record is substantial, not locally reproducible by default
Aleph Alpha reports pre-training completion September 11 after 21 days on 768 Nvidia B200 accelerators, with 38 unplanned interruptions automatically restarted from checkpoints. It lists 392,000 GPU-hours for that stage. The pipeline processed over 200 trillion raw tokens before curation; main pre-training used 20 trillion, mid-training 3.44 trillion and long-context adaptation 200 billion.
Post-training added another set of quantities: 174 billion generated synthetic tokens, a 268 billion-token supervised mix after filtering and combining permissively licensed datasets, and over 1.2 million curated reinforcement-learning tasks. These stage-specific figures should not be confused with unique raw documents or simply added to the main corpus percentages.
Estimated energy for disclosed training phases was 950 MWh including data-center overhead, excluding supervised fine-tuning, reinforcement learning, smaller experiments and some other activity. The company describes screening against a blocklist of over 4.5 million URLs and checks for licensing, lawful sourcing, opt-outs and third-party datasets. Documentation supports scrutiny; it does not prove every copyright or privacy question resolved.
Training took place in Germany and Finland, with a claimed German/European-controlled development pipeline and supply chain. That is broader than storing prompts locally, but still involves foreign chips, software, power and specialist suppliers. Sovereignty means governing remaining dependencies, not eliminating them.
Hardware and long context remain operating constraints
Listed minimum serving configurations include two A100 80 GB accelerators, two H100 SXM5 units, or one H200, B200 or B300. Recommendations include two H100 SXM5 or two H200 for some deployments, alongside newer single-accelerator guidance. The 3.46 billion active-parameter count does not make this an ordinary 3.46 billion-parameter desktop model.
In one company comparison, the 78 billion design handled 18 concurrent requests with approximately 256,000-token contexts on two H100s, versus three for an experimental 123 billion configuration, and decoded 28% faster. That is not a procurement guarantee: precision, batching, software, prompt/output length, memory, communication and service targets change production cost.
Kolibri Base scored 69.8 on RULER at 256,000 tokens and 63.2 at one million. Maximum accepted length does not promise reliable use of every detail. Relevant passages can be buried among repetition, contradictions or malicious instructions. Permission-aware retrieval with traceable citations can be preferable to loading an entire case file, including irrelevant personal information, into one request.
Grounding is promising; benchmarks are mixed
Merlin-Arthur training builds positive document contexts and adversarially incomplete ones, teaching an answer when evidence exists and abstention when it does not. Aleph Alpha reports abstention on 44% of unsupported items in one internal evaluation versus 15% for Kolibri Origin. Its German public-sector proxy rose from 0.54 to 0.75. These are company evaluations, not audited production reliability.
Private customer proxies also cover automotive suppliers, semiconductors, industrial drives and aerospace. They may resemble actual work better than academic tests, but outsiders cannot reproduce them without tasks, prompts, scoring and baselines. Correct and unsupported answers, appropriate and unnecessary refusals, and citation accuracy all matter; a high refusal rate alone can trade usefulness for caution.
Company mathematics results include AIME 2025 scores 96.9 English/87.5 German and AIME 2026 scores 96.0/90.0. GPQA Diamond scores were 84.3 English/81.3 German. Selected comparisons with Mistral, Qwen and Nvidia models support competitiveness, including against models with up to four times the active parameters, not universal superiority.
Tool and agent results are less consistent. BFCL v4 overall was 61.4 versus 67.2 for listed Qwen3.6 35B-A3B; multi-turn was 47.5. TerminalBench 2.1 was 27.7, below listed Qwen3.6, Nemotron 3 Super and dense Qwen3.8. Domain-agent scores were 94.7 telecom, 76.7 airline and 38.1 banking, with other models leading several rows. Prompting, reasoning budgets, translated tests, contamination and scoring can affect comparisons.
Government has several routes, not one champion
Aleph Alpha's F13 assistant work with Baden-Württemberg and a cited rollout intended for 80,000 government-agency users establish access to institutions, not proven Kolibri production use. A named deployment needs its actual workflow, active users, human-review requirements, failure categories and verified time savings after checking outputs. Thousands eligible for a pilot is not the same as daily use.
SAP and OpenAI announced OpenAI for Germany in September 2025, combining OpenAI technology with SAP's Delos Cloud and Microsoft Azure arrangements. That emphasizes domestic operation, contracts and managed support rather than owning the model pipeline. Managed services can supply capacity, monitoring and updates; portable weights give stronger continuity and update control but shift operating responsibilities toward the customer.
The federal KIPITZ platform already supports summarization, translation, drafting, rewriting and conversation, and is planned as cloud-agnostic with application-specific open-source models. No cited announcement selects Kolibri for it. Agentic AI Hub pilots cover planning, permits, meeting transcription and public services; a model's tool calling alone supplies neither their permissions nor a complete agent system.
Alternatives include Mistral, European consortium OpenEuroLLM, the German-English Soofi project on Deutsche Telekom's German Industrial AI Cloud, and open models from Meta, Qwen and Nvidia. Mistral also participates in European government initiatives, including France's deployments and explored Franco-German cooperation with SAP. More replaceable options make sovereignty stronger than dependence on a single national champion.
Independent evaluation offers a further route. Bundesdruckerei's MÖVE began with nine German datasets and seven criteria covering administrative/legal work, security, sustainability and democratic values, with security work involving BSI and Fraunhofer AISEC. A separate study of 39 models from 13 providers found no universal governance leader and energy estimates varying over 60 times. The top two reached only 0.671 accuracy on 4,788 political positions from 64 parties. That study did not evaluate Kolibri directly; it shows why European origin alone cannot establish German institutional competence.
Control must survive deployment and corporate change
The proposed Aleph Alpha–Cohere combination remains subject to regulatory approval. Accounts describe a Berlin/Toronto group of over 1,000 employees, operating under Cohere's name. Kolibri is already complete, while the proposed combination still awaits approval and has not taken effect. More research capacity and enterprise distribution could help, but buyers need clarity on intellectual property, jurisdictions, support, future releases and continuity. Current Apache 2.0 weights remain available on those terms; future openness is a separate question.
The EU GPAI Code of Practice is voluntary guidance to help providers demonstrate relevant AI Act obligations. A signatory's participation, GDPR-oriented design or an open license does not certify every downstream application or training-data question. Drafting, search and eligibility decisions have different risks; deployers still own intended use, affected people, security, human oversight, logging and recourse.
Portable base weights also leave possible lock-in in adapters, retrieval indexes, orchestration and monitoring. Prompt injection, data extraction, poisoned documents and unsafe tool calls require application controls. Reading approved manuals should not grant equipment-command authority, and document assistance should not imply unrestricted email access.
Kolibri's next proof is independent replication and sustained named use: German legal and municipal language, regional variation, inaccessible records, conflicting/outdated policies, revoked permissions, malicious instructions, energy, latency and recovery. Hosted/customized behavior must be compared with the downloadable version. Sovereignty becomes credible when institutions can inspect, operate, replace and govern a reliable service—not when a model's nationality or a million-token headline substitutes for those tests.
Sources and related reading
Kolibri release · German training-data analysis · German reasoning research · OpenAI for Germany · Proposed Cohere combination · Kolibri model card · Primary release discovery record · BF16 weights · MÖVE benchmark · Federal KIPITZ platform · Independent German model-governance study · Knowledge-base access governance · Second release discovery record · October5 announcement · EU GPAI Code of Practice · Third release discovery record