Aleph Alpha's Kolibri: What the New AI Model Means for Businesses
Open weights, a focus on German and a large context window: Kolibri expands the options for enterprise AI. Quality in your own workflow and operating costs remain the deciding factors.
KolibriA new model for deciding where AI should run
Aleph Alpha released Kolibri on 3 October 2026: a German-English language model with open weights under Apache 2.0. Its mixture-of-experts architecture contains 78.1 billion parameters, with approximately 3.46 billion active per token. Official announcement.
For businesses, the choice of operating model is particularly relevant. A company using AI to make internal documents accessible needs to decide where the model runs, how information reaches it and who maintains the service. Downloadable weights provide another option.
This article assesses the release using the manufacturer's documentation as of 3 October 2026. We have not conducted our own practical test of Kolibri for this article. The proposed use cases are our editorial assessment.
The technical details at a glance
| Feature | Manufacturer specification |
|---|---|
| Languages | German and English |
| Maximum context window | 1,048,576 tokens |
| Recommended context | 262,144 tokens |
| FP8 weight memory | approximately 78 GB |
| Example of supported hardware | one H200 or two H100 SXM5 GPUs |
Source: Kolibri product page. The memory figure covers the weights; it does not describe the full requirements of a running service.
Fewer active parameters do not mean little memory
A mixture-of-experts model uses only part of the model for each token. This limits active computation. However, infrastructure planning also needs to account for total size: the remaining experts do not disappear from the model when they are inactive.
Context, concurrent requests and the runtime need additional memory. The active parameter count alone therefore cannot determine a suitable desktop GPU. We explain these relationships in our guide to local LLMs, VRAM and quantization.
A large context window needs a specific purpose
Long documents can matter when analysing manuals, project records or extensive case files. A pilot should answer two questions separately: does the material fit into the context, and can the model reliably find the right information within it?
For internal knowledge search, we would still select relevant passages deliberately. This keeps the basis of an answer verifiable and makes it easier to exclude outdated documents or unauthorized access. A large context does not replace that preparation.
German and grounding deserve their own tests
According to Aleph Alpha, German accounted for 21.3% of pretraining tokens. Kolibri supports four reasoning levels: none, low, medium and high. It was also trained to abstain when the context does not support an answer. Training and capability details.
For a German-language specialist assistant, we would derive three test areas from this. First, does it understand the organization's terminology, abbreviations and sentence structures? Second, how do reasoning settings affect answer quality and waiting time? Third, does it recognize when the documents cannot answer a question?
The last point is straightforward to include in an evaluation. Ask questions about information deliberately absent from the supplied documents. Then assess whether the assistant identifies the gap, asks for clarification or generates a plausible explanation anyway.
What the benchmarks tell us
Aleph Alpha publishes the following results, among others:
| Benchmark | Kolibri | Qwen3.6-35B-A3B |
|---|---|---|
| AIME 2026, German | 90.0% | 84.4% |
| BFCL v4, overall | 61.4% | 67.2% |
Source: Manufacturer benchmarks. These are published manufacturer figures, not our measurements.
The two examples present a mixed picture: Kolibri leads in this mathematics evaluation and trails in the listed tool-calling benchmark. We would not derive a general ranking for a specific application from these results. Solving a mathematical problem correctly and reliably calling the right tool are different capabilities.
A support assistant, for example, needs to locate the relevant case, recognize missing information and return a result in the expected structure. That combination should be evaluated using your own cases. General benchmarks help select candidates; the decision belongs in the actual workflow.
Where we would start a pilot
A suitable starting point would be an internal assistant for a limited collection of German work instructions. The task stays manageable: answer questions, cite passages and identify missing information. The team can review the results before connecting additional workflows.
For this comparison, we would:
- Collect real questions: Include frequent requests, ambiguous wording and cases without a supported answer.
- Set a reference: Compare Kolibri with an existing model under the same conditions.
- Assess quality: Record correct statements, relevant citations and the handling of knowledge gaps.
- Measure operation: Test waiting time, memory and throughput with the expected number of concurrent users.
- Include the effort: Account for infrastructure, maintenance and human correction in the cost per usable result.
The decision to operate a model yourself should cover the entire workflow. Document retrieval, permissions, logging and connected tools matter alongside the answer model. For us, control over AI means being able to deliberately design and verify these parts.
Our assessment
Kolibri is a reason to reassess model selection for German-language business processes. A comparison is particularly interesting where company-controlled infrastructure and internal documents are part of the requirements.
Whether it becomes a useful production service depends on a pilot with clear success criteria. At reamote, we would start with a narrowly defined task and expand once quality, response times and maintenance effort fit together. We discuss how these applications can connect to existing systems on our AI automation page.