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 open-weight model with 78 billion total parameters, 3 billion active parameters and a context window of up to one million tokens. The company says its weights are available under the Apache 2.0 license and presents benchmark results and internal customer-proxy evaluations; independent verification and deployment details are not included in the supplied report.
Aleph Alpha has announced Kolibri, an English-German open-weight model with 78 billion total parameters, of which 3 billion are active, and a context window of up to one million tokens. The company says the full weights can be downloaded from Hugging Face under the Apache 2.0 license, a release it says is intended to let organizations deploy the model in settings where data control and regulatory requirements matter.
Aleph Alpha describes Kolibri as a Mixture-of-Experts Transformer specialized for German, reasoning, mathematics and agentic tasks. The company says it is aimed at mission-critical work in public administration, industry and aerospace, among other customer applications. Its proposed deployment advantage is that organizations can run the model on their own infrastructure rather than sending internal data to third-party inference services; the report does not give hardware requirements or operational cost figures.
The release follows Kolibri Origin, a 30-billion-total-parameter model with 3 billion active parameters and a 65,000-token context window. Aleph Alpha says both models used the same training pipeline, covering data curation, pre-training, post-training and evaluations. The company reports that the pipeline supported hundreds of ablation experiments and stable training that could recover from hardware failures or interrupted data connections without a person stepping in.
Aleph Alpha’s published benchmark table reports Kolibri scores across mathematics, knowledge, coding, long-context and agentic tasks. For example, it lists 96.98 on AIME 2025, 85.9 on LiveCodeBench v6 and 64.5 on LongBench Pro. The company says Kolibri compares favorably with models that have up to four times as many active parameters. Those are vendor-reported results; the supplied material does not describe an independent evaluation or provide enough testing detail to establish how the scores translate to production use.
Why On-Premise Model Access Matters
Kolibri’s release combines downloadable weights with a license that Aleph Alpha identifies as Apache 2.0. For organizations evaluating AI in regulated or sensitive workflows, that offers a route to examine and deploy a model without relying solely on an external inference service. Whether this is practical will depend on computing needs, integration work and the organization’s own security and compliance checks, details not specified in the announcement.
The model’s claimed balance between capability and serving cost is central to Aleph Alpha’s case. Its 3 billion active parameters are presented as a way to make serving more efficient than activating the full 78 billion parameters for each operation. If the company’s comparisons hold under independent and real-world testing, that could make it easier for some organizations to consider locally hosted AI. The announcement does not provide a cost-per-use figure or a direct comparison of total deployment expenses.
Aleph Alpha also frames sovereignty as covering both model development and customer control after release. It says it tracks decisions from data ingestion through evaluation and gives customers deployment freedom and intellectual-property safeguards. Those descriptions are company claims; the report supplied here does not set out a third-party audit, a detailed data provenance record or the specific legal terms beyond the stated license.
high performance AI inference server
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From Kolibri Origin to Release
Aleph Alpha presents Kolibri as a continuation of its model-training work rather than a standalone launch. The earlier Kolibri Origin had the same number of active parameters but fewer total parameters and a substantially shorter context window. The company says iterating on its training pipeline allowed it to run more experiments and release the newer model in less time, though the announcement does not give dates for both releases or a precise development timeline.
For sector-specific testing, Aleph Alpha says it created internal evaluation suites for areas including the German public sector, aviation, manufacturing and automotive work. These evaluations are intended to represent sector language, workflows and edge cases. The company reports score changes on three customer-proxy benchmarks: automotive supplier, 0.72 to 0.99; semiconductors, 0.35 to 0.80; and German public sector, 0.54 to 0.70. The report describes paired synthetic training environments and says customer data was not used for training. It does not provide the scoring methodology or independent validation for these internal results.
“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 Testing and Deployment Details
The supplied announcement does not establish whether Kolibri’s benchmark results have been independently reproduced. It also does not specify the hardware needed to serve the model, throughput under particular workloads, deployment costs, or the full conditions behind the internal customer-proxy scores. Although Aleph Alpha points readers to a technical report for more information, its contents are not included in the source material used here.
The report says customers receive deployment flexibility and intellectual-property safety, but it does not explain the precise legal protections, model update process or limits on commercial use beyond naming the Apache 2.0 license. How well Kolibri performs across individual regulated workflows will depend on testing with those workflows and on each organization’s own technical and compliance requirements.
As an affiliate, we earn on qualifying purchases.
Technical Report and Customer Testing
Aleph Alpha says its technical report contains further details about Kolibri’s development and evaluations. The company has also described a route for customers to assess the model using sector-focused evaluations and synthetic training environments. The next evidence readers will need includes fuller documentation of benchmark methods, hardware and serving costs, license implementation, and results from deployments outside the company’s own reported tests.
For organizations considering Kolibri, the immediate next step is to review the downloadable weights and technical documentation, then test the model against their own data-handling rules and operational tasks. The announcement does not name specific customer deployments or provide a schedule for further releases, so those details remain pending.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Kolibri?
Kolibri is an English-German Mixture-of-Experts model announced by Aleph Alpha. The company says it has 78 billion total parameters, with 3 billion active.
Can organizations download Kolibri’s weights?
Aleph Alpha says the full weights are available on Hugging Face under the Apache 2.0 license. Users should review the applicable license terms and deployment documentation.
What is Kolibri’s context window?
The company says Kolibri supports a context length of up to one million tokens. The announcement does not provide workload-specific performance or hardware requirements for using that maximum.
Has Kolibri’s performance been independently verified?
The source material reports Aleph Alpha’s benchmark and internal evaluation results, but does not identify an independent verification. The figures should be treated as company-reported until external testing is available.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
