Aleph Alpha Launches Kolibri: 78.1B MoE Model Released

Aleph Alpha has introduced Kolibri, a 78.1-billion-parameter Mixture-of-Experts language model designed for English and German text processing. Officially released on October 4, 2026, the model leverages sparse activation to engage only 3.46 billion parameters per forward pass, delivering computational efficiency without sacrificing scale. The release positions Kolibri as a compelling option for enterprises and researchers working with bilingual European language tasks.

Sparse Activation Architecture

Kolibri employs a Mixture-of-Experts architecture, routing each input token through a subset of specialized expert networks rather than activating the full parameter set. This design reduces inference cost and memory overhead while maintaining the representational capacity of a dense 78.1B model. The 3.46B active parameter count per token enables deployment on a single NVIDIA B200 or H200 GPU when using FP8 precision weights, a hardware profile accessible to mid-sized research teams and production environments.

Extended Context and Reasoning Controls

The model supports a one-million-token context window, accommodating long-form documents, codebases, and multi-turn conversations without truncation. Aleph Alpha has integrated per-request reasoning effort controls, allowing users to adjust compute allocation dynamically based on task complexity. This feature differentiates Kolibri from fixed-inference models, offering a dial for balancing speed and thoroughness in real-time applications.

Use cases span legal document analysis, technical support chatbots, and cross-lingual research workflows where extended context retention is critical. The bilingual focus on English and German aligns with European Union regulatory environments and enterprise deployments requiring GDPR-compliant data processing.

Release Details and Availability

Kolibri is distributed under the Apache 2.0 license, enabling commercial use, modification, and redistribution without royalty obligations. The FP8 quantized weights are available for download, reducing storage requirements and accelerating inference on compatible hardware. Aleph Alpha has positioned the model as an open-weight release rather than a fully open-source project, meaning the training code and datasets remain proprietary while the model artifacts are freely accessible.

The company has not disclosed training data composition, compute budget, or benchmark scores in the initial announcement. Early adopters will need to conduct internal evaluations to assess performance on domain-specific tasks such as translation accuracy, reasoning benchmarks, and instruction following.

Competitive Landscape

Kolibri enters a crowded field of open-weight and commercial language models. Recent launches include InclusionAI's Ling 3.0 Flash Fin model and Meta's Muse Spark 1.3 multimodal model, both targeting specialized domains with smaller active parameter counts. Kolibri's bilingual emphasis and European provenance may appeal to organizations prioritizing data sovereignty and regional language support over global multilingual coverage.

The sparse MoE design mirrors trends seen in models like Mixtral and DeepSeek, where selective expert activation delivers cost-performance advantages over dense architectures. However, Kolibri's relatively narrow language support (two languages versus dozens in competing models) suggests a strategic bet on depth over breadth.

What This Means

Aleph Alpha's Kolibri release demonstrates continued momentum in open-weight model development, particularly for non-English language pairs underserved by frontier labs. The combination of Apache 2.0 licensing, single-GPU inference, and million-token context lowers barriers to experimentation and production deployment. Organizations evaluating Kolibri should benchmark against task-specific baselines and consider the trade-offs between bilingual specialization and broader multilingual capabilities offered by alternative models.