TL;DR
Get tech for your team delivered free — and shop member deals
- Fast, free delivery on millions of items
- Access to Prime Big Deal Days deals on October 6–7
- Prime Video, Amazon Music and more included
Aleph Alpha has released Kolibri, an English-German mixture-of-experts model with 78 billion total parameters and 3 billion active parameters. The company says its full weights are downloadable under Apache 2.0 and that the model supports context windows up to 1 million tokens; its performance and deployment claims are based on company-published evaluations.
Aleph Alpha has released Kolibri, an English-German mixture-of-experts model with 78 billion total parameters, of which 3 billion are active, and a context window of up to 1 million tokens. The company says the full weights are available to download from Hugging Face under the Apache 2.0 license, a release that gives organizations an option to run the model under their own deployment arrangements rather than relying only on a hosted service.
Aleph Alpha describes Kolibri as a model for regulated and mission-critical work, including public administration, industrial uses and aerospace. The company says it specialized the model for German-language tasks, reasoning, mathematics and agentic behavior, with the aim of serving workflows where sector-specific language, rules and procedures matter. The report does not independently verify the model’s suitability for any particular deployment.
The release follows Kolibri Origin, an earlier model with 30 billion total parameters, 3 billion active parameters and a 65,000-token context window. Aleph Alpha says both models were developed using its training pipeline, which covers data curation, pre-training, post-training and evaluation. The company says the pipeline supported hundreds of ablation experiments and stable training that could recover from hardware failures or interruptions to data connections.
In its published benchmark results, Aleph Alpha reports competitive scores for Kolibri across mathematics, coding, long-context and agentic tasks, and says it can match models with up to four times as many active parameters on those tasks. These are company-reported comparisons, not independent findings. The report also presents internal customer-proxy results for automotive suppliers, semiconductors and German public administration; those tests are not the same as measured outcomes from customer deployments.
Open Weights for Regulated Deployments
Kolibri’s open-weight release may matter to organizations that need more control over where their AI runs and how data is handled. Aleph Alpha says customers can deploy the model on-premises, keeping internal information away from third-party inference services. The Apache 2.0 license and downloadable weights also give technical teams a basis to examine and adapt the model, subject to the license terms and their own operational requirements.
The company frames sovereignty as both control over how a model is built and the ability to transfer it to customers. It says it can account for the stages from data processing through evaluation and offer customers deployment freedom and intellectual-property protections. Those are meaningful considerations in public-sector and industrial settings, but the report does not establish that every customer will meet its legal, security or compliance requirements simply by using Kolibri.
The model’s mixture-of-experts design is also relevant to serving costs: it has 78 billion total parameters but 3 billion active, according to Aleph Alpha. The company says Kolibri sits on a favorable quality-versus-serving-cost frontier compared with selected models. Actual costs will depend on hardware, workload, configuration and the benchmarks used, so the reported comparison should not be treated as a universal cost guarantee.
As an affiliate, we earn on qualifying purchases.
From Kolibri Origin to Release
Aleph Alpha presents Kolibri as the next step in a continuing model-training effort, rather than a standalone release. Its predecessor, Kolibri Origin, used the same stated number of active parameters but had fewer total parameters and a much shorter context window. The company says the time invested in building and refining its training pipeline helped improve the newer model and shorten the interval between releases.
The announcement is dated March 10, 2026 in the supplied report, which also says the model is being released on the Day of German Reunification. Those details appear inconsistent: German Reunification Day is observed on October 3. The available source does not explain the date discrepancy, so the report’s stated publication date and release-day description cannot both be reconciled from the information provided.
Aleph Alpha says its sector-specific evaluations use internal benchmark suites designed around workflows in fields such as aviation, manufacturing, automotive and the German public sector. It also says paired synthetic training environments let it improve performance without training on customer data. These details describe the company’s approach; the source material does not provide an independent audit of the evaluation design or data practices.
“Kolibri is an English-German Mixture-of-Experts Transformer with 78B total parameters, 3B active.”
— Aleph Alpha
As an affiliate, we earn on qualifying purchases.
Performance and Release Details
The benchmark figures and claims about cost efficiency come from Aleph Alpha’s own report. The supplied material does not identify an independent evaluation or provide enough detail to establish whether the comparisons use identical hardware, serving configurations and testing conditions. Results in production may differ from benchmark scores.
It is also unclear what support, update schedule or long-term maintenance Aleph Alpha will provide for the downloadable weights. The report describes deployment flexibility and intellectual-property safety as company commitments but does not spell out how those protections apply to every use case. The discrepancy between the stated March 10 publication date and the reference to German Reunification Day also remains unexplained.
As an affiliate, we earn on qualifying purchases.
Access, Testing and Deployment
Organizations interested in Kolibri can obtain the weights from Hugging Face under the Apache 2.0 terms, as stated by Aleph Alpha. The company points readers to a technical report for further details on the model and its development. Prospective users will need to review the license and technical documentation, then test performance, operating costs, security and compliance against their own workloads.
Further independent evaluations, deployment examples and clarification of the release-date discrepancy would help establish how the model performs outside Aleph Alpha’s reported tests. The source material does not provide a date for those developments or identify a specific next milestone.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Kolibri?
Kolibri is an English-German mixture-of-experts model released by Aleph Alpha. The company says it has 78 billion total parameters, with 3 billion active.
Can organizations download and run Kolibri?
Aleph Alpha says the full weights are downloadable from Hugging Face under Apache 2.0. The company says the model can be deployed on-premises, but users should check the license and technical requirements for their intended use.
How long is Kolibri’s context window?
Aleph Alpha says Kolibri supports context lengths of up to 1 million tokens. The supplied report does not provide an independent test of that limit.
Are the reported benchmark results independently verified?
Not in the supplied source material. The benchmark scores and comparisons are reported by Aleph Alpha; no independent evaluation is cited there.
Who is Kolibri designed for?
Aleph Alpha says it is aimed at regulated and mission-critical sectors, including public administration, industrials and aerospace, with specialization for German-language work and other capabilities.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
