Skip to content
TrustList
K
AI Model

Kolibri

Aleph Alpha's open-weight English-German mixture-of-experts model for self-hosting, released under Apache 2.0 in October 2026

About Kolibri

Public administration, industry and aerospace are the regulated settings that Aleph Alpha has in mind for Kolibri, an English-German language model released on 3 October 2026. The German AI company presents it as a sovereign, open-weight model that an organisation can run under its own control.

A large model with a small active core

Kolibri uses a mixture-of-experts transformer design: its parameters add up to 78 billion, but only about 3 billion take part in processing any one token. Aleph Alpha says inputs can run as long as 1 million tokens. The reasoning behind the design is cost: with so few parameters working at once, each token needs much less computing than a dense model of similar overall size would demand, a saving that matters to organisations that must serve models on their own infrastructure.

Weights and self-hosting

Hugging Face hosts the full weights, released under Apache 2.0, in an FP8 version built on a BF16 base. The company describes self-hosting with vLLM, and it states no price for a hosted API.

German, reasoning and agents

Aleph Alpha says it specialised the model for German, for reasoning and mathematics, and for agentic use. It also says it validated the training pipeline first on a smaller model, Kolibri Origin, which has 30 billion parameters in total, 3 billion active and a context of 65,000 tokens.

Benchmark comparisons

In the announcement, Kolibri is set beside other post-trained models across English and German benchmarks, and the company positions it at the leading edge when quality is weighed against the cost of serving it. Those results are the company's own evaluations, so readers should treat them as claims to test on their own tasks.

TrustList summary

Kolibri is an English-German language model from Aleph Alpha, the German AI company, released on 3 October 2026. Aleph Alpha positions it as a sovereign model with open weights for regulated sectors, including public administration, industry and aerospace. It is a mixture-of-experts model with 78 billion parameters overall, and roughly 3 billion of them are active for each token.

Main features

  • Long context: supports contexts of up to 1 million tokens.
  • German and reasoning focus: specialised for German, reasoning, mathematics and agentic use, according to the company.
  • Efficient serving: with few parameters active per token, it needs far less compute per token than a dense model of comparable total size.
  • Self-hosting: the full weights are released in FP8 form on a BF16 base, and the company describes self-hosting with vLLM.

Customers and markets

Organisations that must run models on their own infrastructure, including in public administration, industry and aerospace.

Key facts

  • Released: 3 October 2026
  • Licence: Apache 2.0
  • Validation: the training pipeline was first tested on Kolibri Origin, a 30 billion parameter model with a 65,000-token context
  • Benchmarks: the company's own evaluations, not independently checked

Summary written by TrustList from the business's own published information, Nov 2026.

Benchmarks & AI stats

BenchmarkOfficialCommunity avg
No benchmark scores yet. Be the first to add one.

“Official” values are editor-approved and feed the ranking. “Community avg” is the mean of member submissions (shown for transparency; it never affects the ranking until an editor approves a value).

Sign in to add a benchmark score for this model.

Kolibri launches

Categories & features

What the business tells TrustList

Founded
2026 (0 years) evidence sent

Request a demo or quote from Kolibri

Protected by reCAPTCHA - Google Privacy Policy and Terms apply.

By sending, you agree we may share your request and contact details with the provider once you confirm your email.

Reviews

Write the first review of Kolibri

Used it? Your experience helps other buyers decide.

Write a review

Questions & answers

No questions yet. Be the first to ask about Kolibri.

Kolibri alternatives

Similar listings buyers compare against this one.

Google's frontier model, announced in September 2026 and released first to selected cyber defenders

OpenAI's mid-tier GPT-6 model with a 1.05-million token context, image input and multi-agent support in beta

Diffusion language model from Inception for enterprise teams building real-time voice agents, announced in September 2026

Open-weight translation model from Bilibili's Index team, covering 150 languages and following glossary and style instructions

Perplexity's open-weight decision model that picks among given options for routing and checks in AI systems

Mid-tier Anthropic model for everyday work, coding and document creation, with a 1 million token context window

Open-weight agentic coding model from IQuestLab with a 524,288-token context, for teams that self-host

Cloudflare's open-weight decision model that classifies text or images into typed answers with probabilities for AI agents