Kolibri AI: What It Is and What It Costs to Run

Kolibri is an open-weight model created by Aleph Alpha. It is an English-German Mixture-of-Experts (MoE) Large Language Model (LLM) released on 3 October 2026 under the Apache 2.0 license. At Layer3Labs, we build and operate Artificial Intelligence (AI) systems inside other people's businesses.

Aleph Alpha built the model to run on a company's own servers. Because Aleph Alpha published no hosted Application Programming Interface (API) pricing, your expenses depend on computing hardware and engineering staff. Organizations choose this structure when internal records cannot leave company boundaries.

What Kolibri Is and What It Costs

Kolibri is a 78B open-weight model created by Aleph Alpha that is free to download under an Apache 2.0 license. The weights are available on Hugging Face. The Apache 2.0 terms provide a permissive open-source license, and teams should read the full license text on the repository page. Downloading the weights costs zero dollars.

Your actual expenditure comes from hosting hardware and personnel. Aleph Alpha published no hosted API rates on its launch post or homepage. The organization directs teams seeking enterprise deployment and model specialization to contact its sales team. Teams that run the system internally pay for Graphics Processing Unit (GPU) clusters, power, and maintenance engineers.

Aleph Alpha designed the system for on-premise execution. In its launch post, Aleph Alpha compared running a 123B model against this 78B configuration on two NVIDIA H100 cards. Aleph Alpha reported that the 123B alternative handled only three concurrent 256k-token requests, whereas the 78B choice supported 18 queries. Aleph Alpha also reported 28 percent faster decoding.

• Model weights are free to download under an Apache 2.0 license.
• Aleph Alpha published no hosted API pricing for the release.
• Commercial enterprise deployment requires direct outreach to Aleph Alpha sales.
• Hosting expenses depend entirely on internal GPU hardware and engineering staff.

Bilingual Design and Sovereign AI Requirements

Aleph Alpha built Kolibri specifically for bilingual English and German tasks under European standards. The system trained on 768 NVIDIA B200 GPUs in Germany and Finland. It completed 20T pre-training tokens. Mid-training and long-context adaptation brought the total training volume to nearly 24T tokens.

The custom tokenizer contains 128,000 vocabulary tokens balanced for both languages. German material accounts for 21.3 percent of pre-training tokens, with machine translation restricted to 6 percent. Knowledge cutoff dates end at 18 June 2026 for both English and German sources. Technical details appear in the Aleph Alpha tech report.

Aleph Alpha stated that it designed the model around the European Union (EU) AI Act, the General-Purpose AI Code of Practice, and General Data Protection Regulation (GDPR) mandates. Aleph Alpha also claimed complete control across its entire data pipeline from collection to post-training. These statements represent vendor claims rather than legal compliance guarantees for end users. Organizations processing sensitive records must conduct their own legal assessments, as discussed in the GDPR AI compliance guide.

• Trained across 768 NVIDIA B200 processors inside Germany and Finland.
• Pre-training dataset incorporated 21.3 percent native German text.
• Custom tokenizer uses 128,000 vocabulary tokens for bilingual efficiency.
• Aleph Alpha claims full pipeline control, but organizations must verify regulatory standing.

Parameter Scaling and Active Routing in Kolibri AI

Kolibri uses a sparse Mixture-of-Experts architecture containing 78.1B total parameters with only 3.46B parameters active per token. The neural network has 50 layers. Computation distributes across 384 separate expert sub-networks. Only 6 experts activate for any individual token.

A sparse routing system directs incoming text through chosen subsets of parameters instead of calculating the entire network. This mechanism keeps per-token compute demands low while preserving the capacity of a large model. An earlier internal version called Kolibri Origin had 30B total parameters and 3B active parameters with a 65k context window, but Aleph Alpha never released it publicly. The current architecture expands both network scale and document capacity.

Aleph Alpha advertises context support up to 1M tokens, though the maximum trained length is 262,144 tokens. Serving requests beyond 262,144 tokens requires adding configuration flags for max model length and position embeddings during startup. Context beyond trained boundaries requires testing on your own documents to observe degradation. Compare this structure against alternatives in the best open-source LLM guide.

• Architecture contains 78.1B total parameters across 50 layers.
• Active execution routes each token through 6 of 384 expert sub-networks.
• Pre-training natively covers 262,144 tokens, with flags permitting up to 1,048,576 tokens.
• Active compute requirements resemble small models while retaining broader parameter capacity.

Benchmark Claims and Evaluation Realities

Aleph Alpha published internal benchmark results comparing Kolibri against three prominent open-weight models. These metrics come directly from Aleph Alpha without independent verification. On the AIME 2025 mathematics benchmark, Kolibri scored 96.9 against 91.7 for Nemotron 3 Super 120B-A12B, 79.8 for Mistral Small 4 119B-A6B, and 84.6 for Qwen3.6-35B-A3B. The German translation of AIME 2025 yielded 87.5 for Kolibri.

General reasoning and coding evaluations showed competitive metrics. Kolibri scored 84.3 on GPQA diamond and 85.9 on LiveCodeBench v6. On BrowseComp, it registered 29.4, while Nemotron reached 29.1 and Qwen recorded 26.9. However, Kolibri did not lead across every test category. On the AA-Omniscience Index, where higher is better, Kolibri scored -32.8, losing to Mistral Small 4 at -24.0 and Qwen3.6-35B-A3B at -15.3.

On the BFCL v4 tool-calling benchmark, Kolibri scored 61.4, falling behind Qwen3.6-35B-A3B at 67.2. On tau2-bench telecom, Kolibri achieved 94.7, trailing Qwen at 99.1. Aleph Alpha trained the system with an abstention protocol called Merlin-Arthur to encourage the model to state when context lacks necessary answers. The model supports four reasoning effort settings: none, low, medium, and high. Aleph Alpha says higher effort trades cost and latency for answer quality. Check the Hugging Face model card for the request parameter that sets the level. Our suggestion is to start at the lowest reasoning level that passes your internal test set.

• AIME 2025: Kolibri 96.9, Nemotron 3 Super 91.7, Mistral Small 4 79.8, Qwen3.6 84.6.
• GPQA diamond: Kolibri 84.3, Nemotron 3 Super 78.0, Mistral Small 4 74.7, Qwen3.6 83.4.
• AA-Omniscience Index: Kolibri -32.8, lagging Mistral Small 4 (-24.0) and Qwen3.6 (-15.3).
• BFCL v4 function calling: Kolibri 61.4, trailing Qwen3.6 at 67.2.
• Four reasoning effort levels (none, low, medium, high) trade cost and latency for answer quality, per Aleph Alpha.

Enterprise Tasks and Kolibri AI Document Retrieval

Aleph Alpha targets public administration, industrial operations, and aerospace organizations. These categories represent Aleph Alpha targets rather than audited commercial rollouts. The primary task suited for this model is document-grounded question answering over internal corporate records. Organizations retain physical control over their inputs because processing stays inside private servers.

Aleph Alpha built Kolibri for German-language workflows: 21.3 percent of its pre-training tokens are German, and it uses a bilingual German and English tokenizer. Whether that helps on your administrative or legal documents is something only your own test set will show. The Merlin-Arthur abstention protocol assists this work by instructing the model to declare when source documents lack a factual answer.

Serving the model requires specialized software. Aleph Alpha states that Kolibri requires its aleph-alpha-inference package, which provides the Kolibri vLLM plugin. Engineers deploy the environment using the container image at ghcr.io/aleph-alpha/aleph-alpha-inference or by installing the package via pip. The startup command executes vLLM with fp8 kv-cache. It enables auto tool choice and custom parsers. Recommended sampling parameters specify a temperature of 1.0, top_p of 0.97, and top_k of 128. Review agent designs in our guide on self-hosted AI agents for business.

• Targeted by Aleph Alpha for public administration, industrials, and aerospace.
• Document question answering keeps sensitive files within private infrastructure.
• 21.3 percent of pre-training tokens are German, with a bilingual tokenizer.
• Execution requires the aleph-alpha-inference package and the Kolibri vLLM plugin.

Teams That Should Avoid Kolibri

Kolibri is impractical for teams lacking dedicated computing clusters and machine learning engineering capacity. Running a 78B parameter model requires GPU servers, and Aleph Alpha publishes no minimum hardware figure. The launch post documents inference through vLLM only, and no alternative community formats like GGUF or AWQ have been confirmed. Teams seeking simple single-command tools such as Ollama will find no validated setup paths in the launch documentation.

Organizations wanting immediate API access without infrastructure commitments should look elsewhere. Because Aleph Alpha publishes no hosted API price, a test means arranging GPUs yourself or contacting Aleph Alpha sales. If your workflow operates primarily in English and does not face European data residency rules, general alternatives offer simpler deployment. Review the operational guides on how to run Llama locally and how to run Mistral locally for mature alternatives.

Two developments would change this assessment. First, if Aleph Alpha introduces a public pay-per-token API, small teams could experiment without upfront cluster investments. Second, broader runtime integration across standard tools would lower the barrier for non-specialist engineers.

• Requires GPU servers and machine learning engineering expertise.
• Aleph Alpha publishes no hosted API price.
• Tooling relies strictly on the official vLLM plugin with unconfirmed support elsewhere.
• Mature open-weight alternatives exist for workflows without strict German data residency.

Pre-Deployment Verification for Kolibri AI Systems

Teams planning a deployment must validate model performance against realistic corporate files. Begin with document-grounded question answering across internal German-language policies. Evaluate whether the Merlin-Arthur abstention mechanism correctly refuses to answer when documentation is absent. Measuring hallucination rates on your exact source records reveals whether the model fits your quality standard.

Next, verify inference throughput on your target GPU hardware. Measure latency while serving concurrent requests using the official vLLM startup flags. While context extends up to 1M tokens with configuration overrides, test fidelity within the 262,144 trained boundary first. Aleph Alpha's longest trained length is 262,144 tokens, so treat anything longer as untested until your own documents pass.

Deciding between self-hosting and external APIs requires clear cost modeling. A structured AI workflow audit helps organizations evaluate whether private hardware justifies total ongoing operating expenses. Before purchasing hardware, validate your target document tasks with a representative test suite to determine whether kolibri ai justifies the operational commitment.

• Measure retrieval accuracy and abstention performance on internal German documents.
• Benchmark request latency under concurrent load using the official vLLM plugin.
• Validate context retention within the trained 262,144 token boundary before attempting 1M tokens.
• Model hardware capital costs against engineering maintenance before committing to local infrastructure.