# Aleph Alpha releases Kolibri-1 with 1M-token context and tool calling

Digest AI · Generative AI & Models · published 2026-10-03T10:43:51Z · updated 2026-10-03T11:44:42Z

Canonical: https://digestai.news/story/aleph-alpha-releases-kolibri-a-78b-open-weight-llm-for-german-and-engl

## Summary

Aleph Alpha introduced **Kolibri-1**, a mixture-of-experts model optimized for German and English, on October 3, 2026. The model supports explicit reasoning modes, tool calling, and long-context processing up to **1,048,576 tokens** (1M) in principle, though Aleph Alpha recommends contexts of **262,144 tokens** for efficiency. It uses a tokenizer tailored to German morphology and a sliding-window positional encoding to handle extended contexts without scaling issues.

Kolibri-1 was trained in three phases: pre-training on **16,384-token sequences**, mid-training on **65,536-token sequences**, and a final long-context phase on **262,144 tokens**. The model’s training consumed an estimated **9.5×10² MWh** (950 MWh) across all phases, excluding fine-tuning and RL stages. It is designed for integration into conversational assistants, agentic workflows, and document-processing systems where human review is required. Aleph Alpha emphasizes its use for advisory roles, not autonomous decision-making, and includes safeguards against harmful outputs aligned with the **EU AI Act**. The model weights are released under the **Apache 2.0 license**, though Aleph Alpha retains rights to architecture, training methods, and intellectual property.

## Key points

- Kolibri-1 supports **1M-token context** in principle, with recommended use at **262,144 tokens** for efficiency
- Model trained on **20T tokens** pre-training, **3.44T tokens** mid-training, and **201B tokens** long-context adaptation
- Optimized for German/English with **UniBPE tokenizer**, tool calling, and explicit reasoning modes for agentic workflows

## Why it matters

Kolibri-1’s long-context capability and tool-calling features target German/English enterprise use cases like document processing and agentic workflows, filling a niche for non-English AI systems with regulatory compliance safeguards.

## Sources

1. [Aleph-Alpha/Kolibri-1 · Hugging Face - 78B parameters. 3.46B active. Up to 1M tokens of context](https://huggingface.co/Aleph-Alpha/Kolibri-1) (huggingface.co, 2026-10-03, primary source)
2. [Show HN: Germany's new sovereign AI model Kolibri](https://tej.as/blog/aleph-alpha-kolibri) (tej.as, 2026-10-03)

Part of the developing story: [Open-Source LLM Tools Accelerate Real-World AI](https://digestai.news/thread/openarch-launches-pytorch-repo-for-readable-llm-architecture-implementations) (5 stories)

## Cite

Digest AI, "Aleph Alpha releases Kolibri-1 with 1M-token context and tool calling", 3 October 2026, https://digestai.news/story/aleph-alpha-releases-kolibri-a-78b-open-weight-llm-for-german-and-engl

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse
JSON: https://digestai.news/story/aleph-alpha-releases-kolibri-a-78b-open-weight-llm-for-german-and-engl.json