Kolibri: connecting a German runbook to an English incident

The runbook is in German. The incident notes are in English. A support engineer needs an answer that brings them together, not two separate translations. Aleph Alpha’s Kolibri offers a bilingual model to consider for that work: its weights were released on October 3, followed by the October 5 announcement describing document, retrieval and agent uses. Running it in-house gives control over the model portion, while sources and integrations still determine the surrounding workflow.

Before choosing a GPU, put two figures beside one another: 3.46 billion active parameters and about 78 GB of FP8 weights. Parameters are the model’s learned numerical settings; FP8 uses an eight-bit format for most weights; some components retain higher precision. “Active” counts the settings used in one token’s calculation, not all the weights the machine must hold.

The selected experts still belong to a larger model

Kolibri selects groups of calculations, called experts, for each token, a piece of input or output. This mixture-of-experts design does less computation than running every expert each time. The others may be selected for a later token, so they do not vanish from memory when this token passes them by.

The documented minimum setups consequently begin with large accelerators: one H200, B200 or B300, or two A100 80 GB or H100 SXM5 units. They provide room beyond the weights as well. Fitting the model is the first question; how many engineers can get timely answers at once is another. Longer requests, longer replies and concurrent sessions add work and request state.

The context figures also describe different choices. The model card gives a native length of 262,144 tokens and validates extension to 1,048,576. For complex tasks or deployments sensitive to latency and throughput, Aleph Alpha recommends the native length or less. The largest supported request is not the recommended starting point for every job.

Bring the German condition into the English incident

The technical report compares quality and serving costs using synthetic prompts and high-concurrency decoding on eight B200s. Its post-trained throughput comparison uses 16K context and each model’s fastest tested parallel layout. That describes a busy eight-GPU experiment, not the wait for one engineer using one GPU.

Give that engineer a question: “Can I restart the worker now?” In an invented German runbook, “Vor dem Neustart muss die Warteschlange leer sein” means the queue must be empty before restarting. An English incident note says jobs are still waiting. The useful bilingual answer connects the two and explains why the stated condition is not met. Translation alone leaves the engineer to do that work. The storage and recovery analysis similarly connects tested configuration to serving work; an eight-GPU chart cannot settle this one-GPU wait.

For a comparison, give Kolibri and the current model both passages and the same question. Look for the empty-queue condition and the still-waiting jobs in each answer. Does the engineer have to repair the connection? Try representative lengths and concurrent sessions, including the first requests after a cold restart. Completed answers and waiting time describe the job more directly than token rate alone. These passages and trials are hypothetical; no Kolibri result is reported here.

If the work mostly uses other languages, or the available accelerator cannot hold the weights and request state, the efficiency headline does not resolve the mismatch. A smaller model or hosted service may be a more feasible comparison. When the bilingual work and hardware fit, Kolibri is a concrete candidate.

Turn the proposed lookup into the right call

There is one more wrinkle: the English note may be old. The assistant could ask to look up the queue now. A tool-call parser turns its generated call into the tool name and arguments the application can execute; a reasoning parser separates intermediate reasoning from the final answer. Kolibri uses Aleph Alpha’s vLLM plug-in, a supported vLLM version and model-specific parsers for those tasks. The release blog describes four reasoning settings, from none to high. Keep that setting in the comparison because extra reasoning changes the work behind the engineer’s wait.

A fresh queue result supplies another observation, not permission to restart. The model card describes human-reviewed outputs and validated tool results as the intended integration. Its Apache 2.0 grant covers published weights and configuration files, not the other artifacts. Running locally leaves the team responsible both for reviewing the artifacts it uses and for checking the sources, answers and actions around inference.

The answer worth having puts the runbook’s condition beside the evidence about whether it holds, in language the engineer can use. If the queue’s current state is missing, say so. If Kolibri makes that connection with less correction and within the required time, its bilingual focus has earned a place in the comparison. The active-parameter number could not answer the engineer’s question on its own.