Kolibri: A Sovereign Open-Weight Model

TL;DR

Prime Big Deal Days · Oct 6–7Offer from Amazon

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

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

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 1 million tokens. The company says the full weights are available on Hugging Face under the Apache 2.0 license; its performance and sector-specific results are based on company-reported evaluations.

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 1 million tokens. The company says the model’s full weights are downloadable from Hugging Face under the Apache 2.0 license, a release intended to let organizations deploy and inspect the model rather than rely only on a hosted service.

Aleph Alpha describes Kolibri as a Mixture-of-Experts Transformer developed for sovereign, mission-critical work in regulated sectors, including public administration, industry and aerospace. It says the model was specialized for German-language tasks, reasoning, mathematics and agentic behavior, among other customer needs. The company’s stated aim is to balance model capability with serving costs, including deployments on customers’ own infrastructure.

The release follows Kolibri Origin, an earlier model with 30 billion total parameters and 3 billion active parameters and a 65,000-token context window. Aleph Alpha says both models used the company’s training pipeline, covering data curation, experiments, pre-training, post-training and evaluation. It reports that the pipeline supported hundreds of ablation experiments and stable training that could recover from hardware failures or dropped data connections without an operator stepping in.

Aleph Alpha’s published benchmark table reports Kolibri scores across mathematics, knowledge, coding, long-context and agentic tasks. The company says the model matches some models with as much as four times its active parameter count on selected tasks. Results vary by benchmark: for example, its table gives Kolibri a score of 96.98 on AIME 2025 and 85.9 on LiveCodeBench v6. These are vendor-reported results; the supplied material does not establish that the comparisons were independently validated.

Open Weights for Regulated Deployments

Kolibri’s release gives organizations another option for using a language model under their own operational controls. With full weights available under Apache 2.0, users can download the model and choose where to run it, subject to the license and their own technical requirements. Aleph Alpha says on-premise deployment can keep internal information away from third-party inference services, a relevant consideration for public agencies and companies handling sensitive or regulated data.

The company also frames sovereignty as more than where a model runs: it says the approach includes traceability across training and evaluation and customer freedom over deployment. Those are Aleph Alpha’s descriptions of its offering, not a substitute for customers’ own reviews of data provenance, security, legal obligations, operating costs or performance. The release’s practical value will depend on whether organizations can run the model efficiently and whether it performs well on their actual workflows.

Aleph Alpha says it evaluated Kolibri on internal customer-proxy suites for areas such as automotive suppliers, semiconductors and the German public sector. It reports score changes from 0.72 to 0.99 for an automotive-supplier proxy, 0.35 to 0.80 for semiconductors and 0.54 to 0.7 for the public-sector proxy. The company says these suites reflect sector workflows and that paired synthetic training environments were used without training on customer data. The supplied report does not provide enough detail to independently assess those scores or infer real-world return on investment.

From Kolibri Origin to Release

The release is presented as the next stage in Aleph Alpha’s model-training work, rather than as a model built separately from its earlier effort. Kolibri Origin provided an initial test of the training pipeline at 30 billion total parameters and a 65,000-token context length. Aleph Alpha says continued work on the pipeline enabled more experiments and a shorter interval between the two releases, although the source does not specify the exact release dates for Kolibri Origin or the duration of training.

Aleph Alpha says it is releasing Kolibri on the Day of German Reunification. Its report is dated March 10, 2026, but the supplied material does not explain the date reference or provide a separate launch timestamp. The central details are the model’s English-German focus, open-weight availability, Apache 2.0 licensing and stated suitability for controlled deployments.

The report also emphasizes that public benchmarks may not reflect the language, rules and procedures of specialized industries. Aleph Alpha says its internal customer-proxy evaluations were designed to represent those needs. Such tests can offer additional evidence about targeted use cases, but their interpretation depends on disclosed methods, datasets and independent replication.

“Kolibri is an English-German Mixture-of-Experts Transformer with 78B total parameters, 3B active.”

— Aleph Alpha

Independent Tests and Deployment Costs

The source material is Aleph Alpha’s own release report and does not include an independent evaluation of the model. It is not clear from the material how the benchmark tests were configured, whether all competing models were evaluated under identical conditions, or how the internal customer-proxy scores were calculated. The company’s performance comparisons should therefore be read as its reported results, not as independently established rankings.

The announcement also does not detail hardware requirements, actual inference costs, model size on disk, or the operational effort required for on-premise use. Nor does it provide the full information needed to assess training-data provenance, the scope of the stated supply-chain accounting, or how the model handles particular regulated workflows. Those factors will matter to organizations deciding whether open weights translate into a suitable and compliant deployment.

Model Access and Technical Details

Aleph Alpha directs readers to Hugging Face for the model weights and to a technical report for fuller details. Prospective users can examine the release materials, verify the license terms, and test the model against their own requirements. Further information about evaluation methods, infrastructure needs and performance on customer workloads would help clarify how the company’s reported results translate into practice.

The next meaningful evidence will come from documented deployments and evaluations outside Aleph Alpha’s own tests. Until those are available, Kolibri’s reported benchmark standing and suitability for specific regulated uses remain claims to be checked against independent testing and each organization’s operational, legal and security requirements.

Key Questions

What is Kolibri?

Kolibri is Aleph Alpha’s English-German Mixture-of-Experts language model, announced with 78 billion total parameters and 3 billion active parameters.

Can organizations download and run Kolibri themselves?

Aleph Alpha says the full weights are available on Hugging Face under Apache 2.0. The company describes on-premise use as an intended deployment option; specific hardware and operating requirements are not included in the supplied material.

How long is Kolibri’s context window?

Aleph Alpha says Kolibri supports a context length of up to 1 million tokens.

Are Kolibri’s benchmark results independently verified?

The supplied figures come from Aleph Alpha’s report. The source material does not establish independent verification or provide enough methodological detail to confirm the comparisons.

What remains unknown about the release?

The supplied material does not fully specify hardware requirements, deployment costs, evaluation methods or independent real-world results. Organizations will need to test the model against their own workloads and compliance needs.

Source: hn