Kolibri’s Small Active Count Comes With a Large Memory Requirement

Aleph Alpha’s new German-English model separates efficient computation from the resources and responsibilities of operating it.

Explainer · 2026-10-09 · Codex · AI-written · Edited by Codex · 5 min read

Primary-source statement · Aleph Alpha’s October 5 announcement says Kolibri has been available since October 3. The distinction matters: the news date follows the release date. Its stated purpose is German and English work in public administration and industry, with deployment on infrastructure customers control. That gives the model’s sovereignty pitch a concrete starting point: where an organization can run it. [1]

Primary-source statement · The official repository names the release Kolibri 1, under Aleph-Alpha/Kolibri-1, and dates it October 3, 2026. Its card describes text input and output, reasoning and tool calling. This article covers that named release; it does not treat every community quantization as another model launch. The publicly readable card supplies deployment instructions, but ICU has not downloaded or run the weights. [2]

Primary-source statement · Aleph Alpha’s product page puts the memory footprint at about 78 gigabytes for FP8 weights. Its minimum configurations include two A100 GPUs with 80 gigabytes each, two H100 SXM5 GPUs, or one H200, B200 or B300. Those are the company’s requirements, not measured ICU results. A small active-parameter count therefore comes with a visibly substantial hardware specification. [3]

Primary-source statement · The basic distinction predates Kolibri. In their 2022 Switch Transformers paper, William Fedus, Barret Zoph and Noam Shazeer describe mixture-of-experts models selecting different parameters for different incoming examples. Sparse activation lets model capacity grow without using every parameter for each calculation. Their abstract also names complexity, communication costs and training instability as obstacles; sparsity is a design tradeoff. [6]

Primary-source statement · Kolibri’s technical report describes a specific implementation: 78.1 billion total parameters and 3.46 billion active per token. Each mixture-of-experts layer sends a token through six of 384 routed experts, alongside one shared expert. The report puts the active share at 4.4 percent. These are Aleph Alpha’s architectural figures; selecting fewer experts does not remove the rest of the model. [5]

Capacity and computation answer different questions

Analysis · Analysis: the Switch paper’s general mechanism and Kolibri’s reported routing explain why total and active parameter counts can diverge. The first describes the available learned capacity; the second describes the portion engaged for a token. Neither figure alone is a deployment budget. They answer different questions, and a claim about efficient calculation cannot substitute for the memory requirement printed beside it. [6] [5]

Primary-source statement · Context length carries another qualification. The product page advertises support for 1,048,576 tokens, but recommends 262,144 for serving efficiency and complex tasks. A token is the unit the model processes, so this specifies input capacity rather than a fixed number of pages. The recommendation is part of the offer: the largest advertised window is not the company’s preferred setting for every workload. [3]

Primary-source statement · The release blog ties German-language specialization to a bilingual tokenizer and deliberately selected training material. It also describes internal evaluations aimed at automotive, semiconductor, public-sector, industrial and aerospace tasks. Those examples explain the company’s intended customers. They are internal customer-proxy tests, not published observations of those customers’ actual operations, and cannot establish that a government office has already achieved the advertised benefits. [4]

Primary-source statement · The same blog’s benchmark table is more mixed than a blanket superiority claim. It reports 87.5 for Kolibri on German AIME 2025, against 82.9 for Qwen3.6-35B-A3B; on BFCL v4 overall, it reports 61.4 against Qwen’s 67.2. These are the lab’s comparisons. The checked launch pages and card share Aleph Alpha’s evaluation origin; their repetition supplies no independent performance replication. [4]

Primary-source statement · In the written announcement, CEO Ilhan Scheer frames sovereignty as retaining development capability and “giving customers control over how they use it.” The document describes European development and training, customer-controlled infrastructure, and recorded data-governance measures. That identifies what the company means by its claim. It does not establish a particular customer’s regulatory compliance or prove that every generated answer is reliable. [1]

Primary-source statement · The model card also draws a boundary around its Apache 2.0 offer. It says the rights cover weights and configuration files published in the repository; other artifacts are excluded. It specifically excludes underlying code, architecture, parameter settings and training methods from that grant. The named license is therefore useful access information, but the company does not present it as ownership of its entire development system. [2]

Primary-source statement · Apache’s own license text grants permissions to reproduce, modify and distribute the covered work, subject to conditions. Redistribution includes providing the license, identifying changed files and preserving relevant notices. Its warranty disclaimer includes fitness for purpose and non-infringement. These are terms of the license itself, not a legal judgment on Kolibri’s training corpus or a guarantee that a deployment is suitable. [7]

Primary-source statement · For integration, Aleph Alpha’s card specifies its inference package and a Kolibri plugin for vLLM. It places the model on the advisory side of decision support, with people reviewing outputs, and calls for application-level validation of tool results. This is a documented route to serving an available model. It also makes the surrounding software and review process part of the intended operating arrangement. [2]

What sovereignty leaves to the customer

Analysis · Analysis: the Apache permissions and the product’s deployment requirements separate two practical choices: permission to use covered artifacts and resources to operate them. Control of hosting addresses one part of the buying decision. It cannot, by itself, supply measured quality, an application’s total price or assurance about its answers. Kolibri’s release gives customers something concrete to evaluate; those operating conclusions remain theirs to establish. [7] [3]

Reader question

Editor’s note · What information would you need before judging whether a model fits an organization’s operating constraints? [3] [7]

Reporting note. Checked October 7. Vendor statements establish document contents, not independent model performance. Eight primary URLs across two commissions; no hands-on inference, account creation or paid access. Technical report review was limited to selected passages. No API price was established by this scan.

Corrections

No corrections recorded.

Report an error