JOURNAL / 2026.10.04

Kolibri: open weights for German and English work, with context and deployment limits

Aleph Alpha releases Kolibri with reasoning and tool use. Its value lies in being able to evaluate and operate the model under your own control; the advertised million-token context and efficiency require a closer reading.

Aleph Alpha released Kolibri on October 3, a text model focused on German and English, with reasoning and tool calling. The announcement targets public administrations and industries that need to handle sensitive information. Downloadable weights are available, beyond a promise of openness: there are FP8 and BF16 versions, with Apache 2.0 licensing for the published weights and configuration files.

The practical change is the ability to test this combination of capabilities on your own infrastructure and retain a specific version. That gives you control over where a case file is processed and when the model reading it changes. Openness does not automatically cover the data or the entire training system, nor does it make a downstream application compliant by inheritance.

Kolibri has 78.1 billion parameters but activates around 3.46 billion per token through a mixture of experts. That distinction explains part of its economic proposition: each step uses only a fraction of the model. It does not reduce the memory needed to host it in the same proportion. The cards estimate around 78 GB for FP8 weights and 156 GB for BF16, before context memory and other runtime overheads. This is an option for infrastructure with high-memory GPUs, rather than a lightweight download for any computer.

A million tokens are not a million equally useful tokens

The technical report, in section 2.3.4.3 and appendix G, distinguishes training up to 262,144 tokens from tests extending context to 1,048,576. The architecture combines layers that look at a nearby window with others attending to the full context; the latter do not use positional encoding. This allows longer inputs without positional rescaling, but does not automatically preserve quality.

RULER tests examine retrieval of inserted information, variable tracking, word extraction and question answering. They show uneven degradation as length increases. In the final model, enabling reasoning helps question answering at the longest lengths, while disabling it favors retrieval of inserted information. The comparison also uses different output budgets, so it does not isolate the effect of reasoning. The provider itself recommends at most 262,144 tokens for complex tasks and serving efficiency.

For a team analyzing documentation, this suggests starting with representative case files within that range and expanding only when the task warrants it. A file fitting in context does not prove that the system correctly connects its clauses, detects contradictions or recognizes missing evidence.

Efficiency also has a specific frame. Appendix A compares text generation at high concurrency, with synthetic inputs and eight B200 GPUs, searching for each model's best configuration. This is useful for workloads dominated by output generation; it does not by itself measure single-query latency, the cost of reading enormous documents or total service expense. The report also acknowledges evaluating the capabilities of some alternatives in BF16 while measuring their speed in FP8, assuming that lower precision preserves the scores.

I find this opportunity for testable specialization more interesting than a position in a general ranking. The report compares Kolibri with open models such as Qwen and Nemotron and documents improvements over Kolibri Origin, an unpublished internal iteration. But its sector tests are internal proxies for applications, rather than guaranteed customer accuracy rates. Some samples are too small to establish a trend, according to the methodology itself.

A useful first test would be an assistant answering questions about German or English documents and pointing to the passage supporting each answer. I would include questions whose answers are absent, contradictory information and long documents with scattered evidence. I would measure verified answers, correct abstentions, time and memory, with a pinned version of the vLLM inference plugin. We have not run Kolibri here: this is an evaluation proposal, rather than a laboratory result.

Kolibri deserves attention because it expands the options for building under your own control. Its decisive test will be whether that freedom produces reliable work at an affordable cost in a particular language and task. The release lets teams begin answering through their own experiments; it does not settle the question in advance.

Sources

Sources accessed on October 4, 2026.

← Back to journal