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 says it has released Kolibri, an English-German mixture-of-experts model with 78 billion total parameters, 3 billion active parameters and a context window of up to 1 million tokens. The company says the full weights are available under the Apache 2.0 license and presents the model for regulated, mission-critical uses; its performance and sovereignty claims are based on company-reported evaluations.
Aleph Alpha has announced 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 model weights are downloadable from Hugging Face under the Apache 2.0 license, a release that gives organizations the option to run the model in their own environments rather than rely solely on an external inference service.
Aleph Alpha describes Kolibri as a model for regulated and mission-critical work in areas including public administration, industry and aerospace. It says the model was specialized for German, reasoning, mathematics and agent-like behavior, with the goal of performing well on customers’ particular workflows. The company also says its training process did not use customer data for the synthetic training environments associated with its internal evaluations.
The company reports that Kolibri was developed using the training pipeline it had earlier tested with Kolibri Origin, a 30-billion-parameter model with 3 billion active parameters and a 65,000-token context window. Aleph Alpha says the pipeline covered data curation, pre-training, post-training and evaluation, and supported hundreds of ablation experiments. It also says training could continue through hardware failures or interrupted data connections without a person intervening.
In results published by Aleph Alpha, Kolibri scores strongly on selected math, German-language, coding, agentic and long-context benchmarks. The company says its scores across English and German place it on a favorable quality-versus-serving-cost frontier. These are vendor-reported comparisons, not independent validation; the source material does not provide a third-party assessment of the results.
Local Deployment for Regulated Work
Kolibri’s release matters to organizations that want to use language models while keeping sensitive information inside their own infrastructure. Because Aleph Alpha offers downloadable weights, customers can choose on-premise deployment, subject to their hardware, operational and licensing requirements. That option may be relevant to public agencies, manufacturers and aerospace firms with strict data-handling rules or concerns about sending internal material to third-party services.
The combination of a large total model and only 3 billion active parameters is intended to balance capability with inference cost. If the company’s performance and efficiency claims hold up in customers’ environments, the model could offer an alternative to larger systems for selected tasks. The practical outcome will depend on the quality of results on each organization’s data, infrastructure costs, and the work needed to deploy and maintain the model.
Aleph Alpha frames sovereignty as both control over how the model is built and freedom for customers to deploy it. Open weights provide access to model parameters, but do not by themselves establish full control of every supply-chain component or make a system compliant with a particular regulation. Buyers will still need to review documentation, licensing, security, data governance and the model’s behavior in their own use cases.
enterprise AI model deployment hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From Kolibri Origin to Release
Kolibri follows Kolibri Origin, which Aleph Alpha describes as an earlier 30-billion-parameter model with the same 3-billion active-parameter count. Its context window was 65,000 tokens, substantially shorter than the up-to-1-million-token window announced for Kolibri. The company presents the newer release as the result of iterative work on training infrastructure as well as model development.
Aleph Alpha says it created internal evaluation suites for sectors such as the German public sector, aviation, manufacturing and automotive. It reports benchmark-score changes from 0.72 to 0.99 for an automotive-supplier proxy, 0.35 to 0.80 for semiconductors, and 0.54 to 0.70 for a public-sector proxy. These are company-reported internal scores; the supplied report does not state enough detail about scoring scales, sample sizes or independent review to treat them as general measures of real-world performance.
The announcement identifies March 10, 2026, and refers to release on the Day of German Reunification, which falls on October 3. The supplied material does not clarify whether March 10 is the publication date and October 3 is a planned release date, or how those dates relate. Aleph Alpha points readers to a technical report for further detail.
“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.
Independent Results and Release Dates
The announcement’s benchmark tables and efficiency comparisons come from Aleph Alpha’s own report. The supplied material does not include independent replication, detailed test protocols for every comparison, or evidence that the results will transfer to customer deployments. It also does not specify the hardware needed to serve the model, the practical cost of running its long-context features, or how performance changes across different prompt lengths and workloads.
The report says Kolibri is available with full weights, but does not provide details here on model-card contents, training-data disclosure, support arrangements or any usage restrictions beyond the Apache 2.0 terms. The relationship between the March 10, 2026 date and the stated German Reunification Day release is also unclear. Those points should be checked against the linked technical report and the current download listing.
large language model training hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Technical Report and Deployment Checks
Aleph Alpha directs readers to its technical report for details about the model and how it was built. Organizations considering Kolibri can compare that documentation with the downloadable weights and license, then test the model against their own workflows, languages, security requirements and infrastructure budgets.
Further evidence will be needed to determine whether the reported benchmark advantages and internal customer-proxy scores persist in independent tests and production deployments. The supplied announcement does not name an external evaluation, customer deployment, or next scheduled release milestone.
AI model deployment infrastructure
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Kolibri?
Kolibri is an English-German mixture-of-experts language model announced by Aleph Alpha. The company describes it as a model for enterprise and government uses, including regulated sectors.
How large is the model?
Aleph Alpha reports 78 billion total parameters, with 3 billion active parameters. It says the model supports a context window of up to 1 million tokens.
Can organizations run Kolibri themselves?
Aleph Alpha says the full weights are downloadable from Hugging Face under the Apache 2.0 license. That allows organizations to consider self-hosting, although actual deployment depends on their hardware and operational requirements.
Are Kolibri’s benchmark results independently verified?
The results in the supplied announcement are reported by Aleph Alpha. The source material does not describe independent replication or third-party verification.
What is still unclear about the release?
The supplied material does not resolve how the stated March 10, 2026 announcement date relates to the report’s reference to release on German Reunification Day, October 3. It also leaves details about deployment hardware, independent evaluations and customer use cases unanswered.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
