Kolibri’s open weights give buyers control. The memory bill stays.

Downloading a language model can give buyers control without removing the bill for running it. Kolibri’s release puts that trade-off on the table for German-and-English software teams.

Aleph Alpha released Kolibri, its German-and-English language model, on October 3. Its model card says the weights—the learned settings behind its answers—are published under Apache 2.0. [2]

That gives an IT buyer another way to keep a model inside its own systems. It also changes who carries the work. A hypothetical German-language document service could exchange an external provider’s usage charges for its own hardware, maintenance and staff review. Downloading the weights does not make those costs disappear.

The model card limits those license rights to the weights and configuration files published in the repository. Open weights should not be confused with permission to use every part of the developer’s technology. [2]

Aleph Alpha lists 78 billion total parameters, with about 3.46 billion active per token. Parameters are the model’s learned numerical settings; tokens are the pieces of text it processes. [2]

The model card puts the memory footprint of its FP8 weights at approximately 78 GB. FP8 is a compact numerical format for storing those settings. [2]

The small active portion explains the design’s appeal: a larger pool of settings need not all do computation for every text fragment. But buyers must budget for the model they store, not just the portion active at a moment. The active-parameter count alone cannot tell them the hardware bill or cost per useful answer.

For serving efficiency and complex tasks, Aleph Alpha recommends contexts of at most 262,144 tokens. Context is the text available to the model while it works on a request. [2]

For a hypothetical team searching long German-language files, the practical question is how many complete jobs it can finish at the required response time. A demonstration with one request does not price a queue of requests. Test the actual document lengths and simultaneous users before sizing the service.

Aleph Alpha describes the intended systems as ones in which a person reviews the model’s output before it is acted on. That is a material boundary for its advertised reasoning and tool use. [2]

Aleph Alpha says its benchmarks were run using its own evaluation software. The reported results are developer-run comparisons, rather than independent measurements of customers’ operating costs. [3]

The strongest counterargument is that self-hosting might transfer work rather than remove it. A small workload could be cheaper with a hosted service once engineering time and unused hardware are counted. Better control over where a model runs does not establish lower costs, correct answers or compliance for a particular application.

The next useful result would be a like-for-like pilot reporting cost per accepted answer on the same German-language documents. Count hardware time, staff corrections and ongoing maintenance. Report response time under the same load and the share of answers needing correction. If review work or delays erase any computing advantage, the savings case weakens. The release expands buyer choice; that comparison will show whether the choice pays off.