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 support for context lengths up to 1 million tokens. The company says the model’s full weights are available under the Apache 2.0 license; its performance claims are based on company-reported benchmarks, and independent verification is not established in the supplied material.
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 weights are downloadable from Hugging Face under the Apache 2.0 license, making the model available for organizations seeking to run AI systems under their own deployment arrangements.
Aleph Alpha describes Kolibri as a model for mission-critical work in regulated sectors, including public administration, industrials and aerospace. It says development focused on German-language performance, reasoning, mathematics and agentic behavior, as well as capabilities requested by customers. The company presents local or on-premises deployment as a way for organizations to process internal data without sending it to third-party inference services.
The model uses a mixture-of-experts architecture: 78 billion parameters are present in total, while 3 billion are active during use, according to Aleph Alpha. The company says this design balances model capability with serving costs. Its report describes Kolibri as lying on the quality-versus-serving-cost Pareto frontier in English and German, based on its comparisons with selected models and benchmark results.
Aleph Alpha reports that Kolibri performs competitively across math, coding, knowledge, long-context and agentic tasks. In its published table, for example, it gives Kolibri scores of 96.98 on AIME 2025 and 85.9 on LiveCodeBench v6. These are company-reported benchmark results, not independent findings established by the material provided. The company also says it created internal evaluation suites for customer-related areas such as automotive suppliers, semiconductors and the German public sector; it reports scores rising from 0.72 to 0.99, 0.35 to 0.80 and 0.54 to 0.70, respectively.
Why Local Deployment Matters
Kolibri’s release gives organizations another option for deploying a downloadable model with published weights rather than relying solely on hosted inference services. For public agencies and companies handling sensitive information, the ability to operate a model within their own infrastructure may affect data handling, procurement and regulatory compliance. Those benefits depend on the customer’s deployment and governance choices; the model’s availability alone does not establish that every use meets a particular legal or security requirement.
The stated 3-billion active parameter count may also matter to operators weighing capability against the cost of running a large model. Aleph Alpha argues that Kolibri compares favorably with larger-active-parameter models on its selected evaluations. Readers should treat that as the vendor’s assessment: benchmark scores do not by themselves establish performance on a customer’s specific workflows, hardware or production workloads.
Open-weight access and the Apache 2.0 license provide room for organizations to inspect and deploy the released model within the license terms. They do not, by themselves, make the full training data or every element of the model’s development public. Aleph Alpha’s broader claims about supply-chain integrity and transparency therefore matter to buyers, but the details need to be checked against the company’s technical report and the organization’s own requirements.
As an affiliate, we earn on qualifying purchases.
From Kolibri Origin to Release
Aleph Alpha says Kolibri followed Kolibri Origin, an earlier model with 30 billion total parameters, 3 billion active parameters and a 65,000-token context window. The company describes both models as products of a training pipeline it developed for data ingestion and curation, pre-training, post-training and evaluation. It says the pipeline supported hundreds of ablation experiments and stable training that could recover from hardware failures or dropped data connections without human intervention.
The announcement identifies German Reunification Day as the intended release occasion, while the supplied report is dated March 10, 2026. The material does not explain the relationship between those dates, so the publication date and stated release occasion should not be treated as interchangeable. Aleph Alpha directs readers to a separate technical report for more detail about how Kolibri was built and evaluated.
The company says it developed internal, sector-focused evaluations and paired synthetic training environments to improve performance without training on customer data. Those statements describe Aleph Alpha’s process; the supplied source does not provide an independent audit of the data pipeline or the internal test suites.
“Kolibri is an English-German Mixture-of-Experts Transformer with 78B total parameters, 3B active.”
— Aleph Alpha
Benchmark Scope and Release Details
The supplied report does not include independent replication of the benchmark results or enough detail here to assess every evaluation setup, hardware configuration and comparison condition. Aleph Alpha says readers can consult its technical report, but the supplied material does not reproduce that report’s full methods. The results should therefore be read as company-reported measurements, not as independently confirmed rankings.
It is also unclear from the provided information how the model’s 1-million-token context limit performs across different tasks, how much compute is required for typical deployments, and what throughput customers should expect on particular hardware. The company’s statements about transparency, supply-chain integrity and intellectual-property safety are not accompanied here by third-party audits or detailed evidence. The stated German Reunification Day release occasion also does not align clearly with the report’s March 10, 2026 date.
Technical Report and Deployment Tests
Aleph Alpha points readers to a technical report for further information on training and evaluation. Organizations considering Kolibri can compare those methods and licensing terms with their own requirements, then test the model on representative workloads and infrastructure. The supplied source does not give a schedule for later releases, independent evaluations or additional deployment documentation.
Further reporting can establish whether outside researchers reproduce the published benchmark results, how the model performs in customer environments, and what practical costs and safeguards are involved in on-premises operation. Until those details are available, the confirmed development is the model’s announced release and the company’s stated specifications and access terms.
Key Questions
What is Kolibri?
Kolibri is an English-German mixture-of-experts language model announced by Aleph Alpha. The company lists 78 billion total parameters and 3 billion active parameters.
How long a context does Kolibri support?
Aleph Alpha says Kolibri supports context lengths of up to 1 million tokens. The supplied material does not give independent testing of performance at that maximum length.
Can organizations download and run the model?
Aleph Alpha says the full weights are available on Hugging Face under the Apache 2.0 license. The company describes on-premises deployment as an intended use, but deployment requirements will depend on an organization’s hardware and operating setup.
Are Kolibri’s benchmark results independently verified?
The supplied report presents results from Aleph Alpha and does not establish independent verification. The company refers readers to a technical report for further methodology.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
