Aleph Alpha Open-Sources Kolibri: A Deep Dive into the 78B-Parameter MoE Model

Kolibri Arrives: A New Open-Source LLM Option from Europe

German AI company Aleph Alpha has released its next-generation open-source model, Kolibri, under the Apache 2.0 license, targeting bilingual English and German use cases. The model's most striking feature is its Mixture of Experts (MoE) architecture: 78 billion total parameters, yet only roughly 3 billion are activated per inference pass — all while supporting a context window of up to 1 million tokens.

For European enterprises and developers, Kolibri's significance goes beyond raw performance metrics. It offers a compliant, self-hostable, and German-optimized open-source option. In the current LLM ecosystem, high-quality open-source models deeply optimized for German are rare — and Kolibri fills that gap.

MoE Architecture: What 78B Total / 3B Active Parameters Actually Means

Mixture of Experts (MoE) is one of the most important architectural directions in recent large model development. By adopting this design, Kolibri retains the knowledge capacity that comes with 78 billion total parameters, while activating only around 3 billion during each forward pass.

The core value of this design lies in decoupling model capacity from inference cost. In traditional dense models, more parameters means slower inference and higher cost. MoE addresses this through a routing mechanism that sends each token through only a small subset of expert networks. This means Kolibri's inference compute is closer to that of a 3B-scale model, while still drawing on the knowledge base of a 78B model.

For deployment teams, this is a highly practical trade-off: no need for top-tier hardware to run a full dense large model, yet you can still achieve near-large-model performance on targeted tasks.

1 Million Token Context Window

Kolibri supports an ultra-long context of up to 1M (one million) tokens. This capability is especially critical for handling long documents, codebase analysis, legal contract review, and complex multi-turn conversations. Combined with its English-German bilingual positioning, Kolibri has a natural edge in European enterprise document processing workflows.

MoE architecture was first proposed by Google in 2017 and has since been validated at scale in models like Mixtral and DeepSeek^Ad. Its core component is the Router, which dynamically selects which "expert" sub-networks (typically variants of feed-forward layers) to activate for each token. In typical implementations, the model selects the Top-2 experts from dozens of candidates per token, leaving the rest idle during that inference step. The direct result: during training, total parameter count can scale up to increase knowledge capacity, while inference FLOPs don't grow proportionally with total parameters. One important caveat — MoE's "sparse activation" advantage is primarily about compute. Memory usage is not reduced — all expert weights still need to reside in VRAM. This is exactly why deploying a 78B MoE model still demands substantial GPU memory.

How to Run It: Deploying with vLLM

According to the official release guide, Kolibri is recommended for inference deployment via vLLM. vLLM is a leading high-performance LLM inference framework that supports optimizations like PagedAttention, enabling efficient VRAM management under long-context workloads.

For teams planning to self-host, the combination of MoE architecture and vLLM means production-grade deployment is achievable with relatively manageable hardware resources. That said, even though only 3B parameters are activated at a time, the full 78B parameter set still needs to be loaded into VRAM — so memory capacity remains the primary hardware constraint.

vLLM was open-sourced by UC Berkeley in 2023. Its core innovation is PagedAttention — borrowing the concept of virtual memory paging from operating systems to divide the KV Cache (key-value cache) into non-contiguous memory blocks allocated on demand. This dramatically reduces memory fragmentation and improves concurrent throughput. When handling million-token ultra-long contexts, KV Cache memory consumption grows linearly with sequence length, making PagedAttention especially critical for this use case. vLLM also supports Continuous Batching, Tensor Parallelism, and more, making it one of the go-to choices for deploying MoE models in production. For a model like Kolibri — which requires sharding all 78B parameters across multiple GPUs — vLLM's tensor parallelism support distributes model weights across cards, lowering the per-GPU memory threshold.

Who Should Consider Using Kolibri

Based on the model's positioning, Kolibri is best suited for the following user groups:

• Enterprises with German-language business needs: German-optimized open-source models are scarce. Kolibri gives German-speaking organizations a high-quality, self-hostable option.
• Organizations prioritizing compliance and data sovereignty: The Apache 2.0 license permits commercial use and derivative works. Self-hosting also makes it easier to meet Europe's stringent data compliance requirements.
• Teams optimizing for inference efficiency: The MoE architecture offers a strong balance between cost and capability — ideal for scenarios where inference cost needs to be controlled without sacrificing quality.

Conversely, if your workload is primarily English-only and you already have a mature closed-source API solution in place, Kolibri's bilingual advantage may not add enough value to justify the switch.

Closing Thoughts

Kolibri represents a meaningful step forward for open-source AI out of Europe. With its Apache 2.0 openness, efficient MoE architecture, million-token context support, and English-German bilingual capabilities, it delivers differentiated value for specific markets. For teams focused on data sovereignty, regulatory compliance, and German-language support, this model is well worth including in your technical evaluation process.

(Note: This article is based on the official release guide. For specific benchmark data and deployment commands, please refer to the official documentation.)