Aleph Alpha Kolibri Review for UK SMEs

Rohan Morris from AI Advisers · 6 October 2026

Aleph Alpha is a German AI company, and Kolibri is its open-weight language model, released 3 October 2026. It has 78.1 billion total parameters, activates about 3.46 billion per token, handles German and English, and is licensed under Apache 2.0. Aleph Alpha says it was built with the EU AI Act in mind.

Aleph Alpha Kolibri at a glance

• Released: 3 October 2026, according to Aleph Alpha's own announcement.
• Size: 78.1 billion total parameters, with about 3.46 billion active for each token.
• Languages: English and German. German makes up 21.3% of pre-training tokens.
• Licence: Apache 2.0, which allows the weights to be downloaded, run and modified.
• Context: Up to 1 million tokens.

What is Aleph Alpha Kolibri?

Kolibri is a mixture-of-experts model. It holds many specialist sub-networks but switches on only a few for each word. That keeps running costs well below a conventional model of the same total size. Aleph Alpha says it trained Kolibri on 768 B200 GPUs and positions it for "sovereign, mission-critical work" in public administration, industrials and aerospace.

Aleph Alpha Kolibri benchmarks: how does it compare?

The table below uses the scores in Aleph Alpha's announcement. It compares Kolibri with three other open-weight models.

| Benchmark | Kolibri | Nemotron 3 Super | Qwen3.6-35B | Mistral Small 4 | |---|---|---|---|---| | AIME 2025 (maths) | 96.9 | 91.7 | 84.6 | 79.8 | | GPQA Diamond (science knowledge) | 84.3 | 78.0 | 83.4 | 74.7 | | HumanEval+ (coding) | 92.7 | 94.7 | 92.8 | 92.8 |

Kolibri leads this table on maths and science knowledge. It is last on HumanEval+, but only by 0.1 to 2.0 points.

Three caveats apply:

• These are Aleph Alpha's own figures. Run your own tests on your own tasks before relying on them.
• The comparison set is dated. An independent review of Kolibri notes: "All three comparison models date from the spring. Since then, Chinese labs in particular have released considerably stronger open-weight models."
• Kolibri does not win everywhere. The same review says Qwen3.6 outperforms Kolibri on agentic tasks and long-context work.

Is Aleph Alpha Kolibri compliant with the EU AI Act?

No model is "approved" under the EU AI Act. The Act sets duties for providers and for businesses that use AI. It does not give a model an approval stamp. The most relevant signal is the voluntary General-Purpose AI Code of Practice.

• Code of Practice: The European Commission describes the Code as "a voluntary tool, prepared by independent experts in a multi-stakeholder process." Aleph Alpha is listed as a signatory.
• Timeline: Obligations for general-purpose AI model providers applied from 2 August 2025. The Commission's enforcement powers, including fines, apply from 2 August 2026.
• Aleph Alpha's claim: The announcement says Kolibri was "built with the EU AI Act, the General-Purpose AI Code of Practice and the GDPR in mind from the ground up." That is a design intention, not a certification.

For a UK business, a provider that has signed the Code can ease your due diligence. It does not remove your own duties. Those depend on how you use the model. Read what the EU AI Act means for UK small businesses in 2026 and our EU AI Act compliance support before you deploy any model with EU-facing output. This is general information, not legal advice.

Can you run Aleph Alpha Kolibri locally?

Yes, with the right hardware. The weights are downloadable under Apache 2.0, so you can host them on your own servers or a private cloud. The model is large, though. The announcement describes serving it on two H100 GPUs, which handled 18 concurrent requests at 256,000-token inputs. A laptop or an office PC will not run it. Most UK small businesses would rent a private GPU server or use a managed host they trust.

What is Aleph Alpha Kolibri best for?

Kolibri suits businesses that handle sensitive documents and want the data to stay in their own environment. It fits best with:

• Document-heavy work: policy questions over internal files, first drafts of contract clauses, and summarising reports.
• German and English content: the model was built for both languages.
• Regulated sectors: the announcement targets public administration, industrials and aerospace.

It is a weaker fit for quick prototypes, where a hosted model is cheaper to start. It is also a weaker fit for agentic tool use and long-context tasks, where the review above rates Qwen3.6 higher.

Our advice is to start with a hosted model and move to self-hosting only once you have a clear data-residency need and enough volume to justify the hardware. If you are unsure which tasks justify that cost, an AI readiness audit will map them. For the workflows around the model, see our guide to workflow automation for small businesses.

Who is behind Aleph Alpha, and what changes with Cohere?

In April 2026, Canadian firm Cohere agreed to acquire Aleph Alpha. The combined company was valued at about $20 billion, with $600 million from the Schwarz Group. Futurum's analysis says Aleph Alpha's PhariaAI platform becomes the deployment layer alongside Cohere's models. Canada's AI Minister framed the goal as enabling "governments and companies have an option between the hyperscalers and the hegemon."

The deal was reported as awaiting regulatory approval in Germany, the EU and Canada at the time of writing. Check the current status before relying on it.

The Apache 2.0 licence applies to the released weights. Still, check the licence terms and support commitments with any vendor before you build on them.

Aleph Alpha Kolibri for UK SMEs: the takeaway

Kolibri gives UK small businesses a European-built, open-weight option with strong maths and science scores on Aleph Alpha's own tests. It is not an EU-approved model, and it is heavy to run. Test it on your own documents before committing, and treat the EU alignment as a starting point for your own compliance work, not a replacement for it.

Aleph Alpha and EU AI Act support for UK SMEs

AI Advisers is based in Milton Keynes and works with small businesses across Milton Keynes, Luton, Bedford, Northampton and the wider UK to choose, test and govern AI tools that fit EU AI Act duties. Start with an AI readiness audit, or explore our EU AI Act compliance support and AI literacy training for your team.

Frequently asked questions

What is Aleph Alpha Kolibri?

Kolibri is an open-weight German and English language model from Aleph Alpha, released on 3 October 2026. It has 78.1 billion total parameters and is licensed under Apache 2.0, so its weights can be downloaded and run on your own hardware.

Is Aleph Alpha Kolibri approved under the EU AI Act?

No. The EU AI Act does not approve individual models. Aleph Alpha is a signatory to the voluntary General-Purpose AI Code of Practice, and it says it built Kolibri with the Act in mind. Your own use of any model still needs its own compliance review.

Can a UK small business run Aleph Alpha Kolibri locally?

Yes, but not on a standard office computer. The announcement describes serving it on two H100 GPUs, which are data-centre hardware. Most UK small businesses would rent a private GPU server or use a managed host.

How does Aleph Alpha Kolibri compare with other open models?

Kolibri scores 96.9 on AIME 2025 and 84.3 on GPQA Diamond, the highest of the four models shown. It scores 92.7 on HumanEval+, the lowest. These figures are self-reported, so test the model on your own tasks.

Is Aleph Alpha owned by Cohere?

Not yet. Cohere agreed in April 2026 to combine with Aleph Alpha, and the deal was reported as awaiting regulatory approval. Check the current status before relying on it.

Which open-weight AI model should UK businesses use?

It depends on your use case. Kolibri is strong for document-heavy, regulated work in English and German, with a focus on data sovereignty. Qwen3.6 scores higher on agentic tasks and long-context work. Mistral Small 4 is lighter to run. If you are unsure which model fits your workflows, an AI readiness audit will map your tasks to the right tools.

Key terms

• Open-weight: The trained model parameters are published, so you can download and run the model yourself.
• Mixture-of-experts (MoE): A model design that activates only some of its sub-networks for each token, which lowers running costs.
• Active parameters: The share of a model's total parameters used for each token.
• Apache 2.0: A permissive open-source licence that allows commercial use, modification and redistribution.
• General-Purpose AI Code of Practice: A voluntary EU framework that helps providers of general-purpose AI models show compliance with the AI Act.