German artificial intelligence developer Note has released “Kolibri,” a new open model optimized for English and German language processing. Announced on October 3, 2026, the model features a Mixture of Experts (MoE) architecture with 78 billion total parameters, activating approximately 3 billion parameters per token. The release utilizes the permissive Apache 2.0 license, allowing organizations to download weights directly from Hugging Face for secure deployment.
The system is specifically engineered for on-premises infrastructure, targeting regulated industries such as public institutions, manufacturing, and aerospace where data privacy and sovereignty are paramount. According to technical specifications outlined in the release, Kolibri supports a massive context window of up to 1 million tokens, accommodating extensive document analysis and long-form technical manuals.
Training Focus and Reliability Guardrails
Developers designed Kolibri to address common reliability challenges in enterprise deployment. Approximately 21.3 percent of the training corpus is composed of German language data, enhancing regional nuance and compliance capabilities. Furthermore, the model has been explicitly trained to decline requests or output a “I don’t know” response when verified information is absent from provided source documents, mitigating the risk of hallucinations in professional environments.
In terms of performance benchmarks, the releasing organization claims Kolibri achieves the highest quality tier within its speed and cost category, competing effectively against models possessing up to four times its active parameter count. However, these efficiency metrics rely entirely on internal measurements by the developer, and independent third-party verification remains unconfirmed.
Deployment Implications and Next Steps
The availability of an open-weights model tailored for European regulatory standards provides an alternative for enterprises seeking sovereign cloud and local hardware options. Organizations evaluating the software will need to conduct rigorous local benchmarking to validate the developer’s speed and cost claims against specific enterprise workloads. As adoption proceeds, industry stakeholders will monitor how effectively the model’s hallucination-reduction guardrails perform in live operational settings across manufacturing and public administration.