Aleph Alpha Releases Kolibri Mixture of Experts Model for Secure Enterprise RAG

German AI company Aleph Alpha has introduced Kolibri, a new open-weight language model designed for organizations that want more control over their data and AI infrastructure. The model mainly targets regulated sectors, where sensitive information cannot always be sent to external cloud services.

Instead of relying entirely on an external model provider, Kolibri can be hosted on infrastructure controlled by the organization itself. Its weights are released under the Apache 2.0 license, giving companies considerable freedom to deploy and adapt the model for their own environment.

Cheaper per Token, But Still Hungry for Hardware

Kolibri uses a Mixture of Experts (MoE) architecture, routing requests through specialized sub-networks rather than activating the entire neural network for every token processed.

The model comprises 78.1 billion total parameters, with 3.46 billion parameters active per token.
This reduces the amount of computation required during inference. However, the full model weights still need to be stored and distributed across the deployment hardware, so Kolibri remains demanding to host despite its relatively low compute cost per token.

Retrieval and Tool Use

Kolibri is post-trained to work in retrieval-augmented environments, where it can access external search and document tools. During training, the model is exposed to several retrieval methods, including vector search, BM25 keyword search, hybrid search, exact pattern matching, and document-reading tools. Aleph Alpha also varies the tools and retrieval settings between training episodes, so the model does not become dependent on a single setup.

This is useful in enterprise environments because different questions require different search strategies: a broad conceptual question may benefit from semantic search, while an exact part number, serial number, or legal reference may require keyword- or pattern-based search.

Kolibri is also trained to abstain when the retrieved information does not provide enough evidence, rather than generating an unsupported answer. As the report notes:

"The effectiveness of any RAG pipeline is bounded by the quality of the underlying knowledge base. If internal documentation is unstructured, outdated, or poorly indexed, even advanced retrieval and routing methods will struggle to return reliable answers" (garbage in, garbage out).

Synthetic Training Pipeline and Benchmark Performance

Because confidential customer data may be restricted, unavailable, or unsuitable for training, Aleph Alpha makes extensive use of synthetic data during post-training. External models — including Qwen3-32B, GLM-5.2, and GLM-5.3 — are used to generate and refine training examples for areas such as long-context reasoning, retrieval, and tool use.

Training then proceeds through two main phases:

Supervised Fine-Tuning (SFT): Teacher models generate structured training examples that show Kolibri how to handle tasks such as retrieval, reasoning, and tool use.
Reinforcement Learning (RL): The model improves through interaction with different training environments. For some tasks, LLM-based judges are used to evaluate responses more flexibly than simple exact-string matching.

One notable detail: Qwen3 is used both to generate parts of Kolibri's synthetic training data and later appears as a benchmark comparison model. This is not inherently problematic, but it is worth keeping in mind when interpreting the results.

In internal evaluations across several RAG benchmarks — including specialized domain workflows in aerospace, semiconductors, German public administration, automotive manufacturing, and industrial engineering — Aleph Alpha reports strong results for Kolibri. According to the company's own evaluation, Kolibri achieved an average accuracy score of 75.9%, compared with 71.1% for Qwen3-35B-A3B and 68.4% for Nemotron 3 Super.

These results are promising, but they should be interpreted carefully: the benchmark setup and reported scores come from Aleph Alpha itself, so independent third-party testing is still needed to confirm how well Kolibri performs in real-world enterprise environments.