Aleph Alpha Releases Kolibri-1 Open-Weights Model Under Apache 2.0
The German AI lab has published the full weights for its 78-billion-parameter bilingual mixture-of-experts model under a permissive open-source license.
Key takeaways
- Aleph Alpha released the weights for Kolibri-1 under the Apache 2.0 license on Hugging Face.
- The 78.1B parameter mixture-of-experts model activates only 3.46B parameters per token across 50 layers.
- Kolibri-1 supports up to 1,048,576 context tokens and is built specifically for German and English workloads.
- The model was trained on 768 Nvidia B200 GPUs across German and Finnish data centers.
Aleph Alpha has released the full weights and architecture configuration for its Kolibri-1 open-weights model on Hugging Face under the permissive Apache 2.0 license. Announced on October 3, 2026, the release makes the German AI developer's 78.1-billion-parameter bilingual system accessible to developers and enterprises aiming to run sovereign infrastructure on their own hardware according to the official Hugging Face repository.
The release represents a shift toward open developer access for Aleph Alpha, which historically tailored its deployments directly to enterprise and European public sector clients, as reported by Startup Fortune. By releasing the model under Apache 2.0, downstream users can self-host, modify, and integrate the system into commercial workflows without licensing fees or copyleft restrictions.
Sparse Mixture-of-Experts Architecture and Specs
Kolibri-1 is structured as a sparse mixture-of-experts (MoE) transformer comprising 50 layers. While the entire model holds 78.1 billion parameters in memory, the system routes each token through only six routed experts out of 384 per layer, plus one shared expert that processes every token. This sparse design keeps compute demand down to roughly 3.46 billion active parameters per token, according to details in the Kolibri-1 Model Card.
To balance memory and context efficiency, 40 of the model's 50 layers employ sliding-window attention constrained to the previous 512 tokens, while every fifth layer executes full attention across the entire context window. Kolibri-1 was natively trained up to 262,144 tokens and validated for context lengths of up to 1,048,576 tokens. Aleph Alpha notes that because rotary positional encodings are confined to the local sliding-window layers, the model extends to longer sequences without custom position scaling.
Running Kolibri-1 requires storing the complete 78.1-billion-parameter model in memory, demanding roughly 78 GB of GPU memory for weights stored in FP8 precision (float8_e4m3fn). According to the model card, minimum hardware requirements include two Nvidia A100 (80 GB) or H100 SXM5 accelerators, or a single Nvidia H200, B200, or B300 GPU.
Bilingual Specialization and Training
Unlike broad multilingual models, Aleph Alpha focused Kolibri-1 exclusively on German and English. As detailed in Aleph Alpha's launch announcement, the model was pre-trained on 20 trillion tokens, with organic German accounting for approximately 21.3% to 23.9% of the dataset, English representing 62.5%, and code comprising 13.6%. The company intentionally limited translated material to 6% of the corpus to avoid unnatural syntactic patterns.
To support complex German compound words, the engineering team built a 128,000-vocabulary bilingual tokenizer using a custom algorithm called UniBPE. An independent analysis by engineer Tejas Kumar testing the German Basic Law showed that Kolibri-1 encoded the constitutional text in 35,190 tokens, compared to 41,482 tokens required by OpenAI's GPT-5 tokenizer.
Pre-training was conducted on 768 Nvidia B200 GPUs (96 HGX nodes) over 21 days across data centers in Germany and Finland, consuming an estimated 950 MWh of energy under German and European legal compliance frameworks.
Abstention and Controllable Reasoning Features
Kolibri-1 includes an explicit reasoning mode alongside native tool-calling capabilities. Developers can toggle the reasoning_effort parameter across four levels—none, low, medium, or high—allowing applications to trade compute latency against depth of thought, as outlined in technical documentation cited by Superpower Daily.
To curb hallucinations in enterprise retrieval-augmented generation (RAG) tasks, Aleph Alpha trained Kolibri-1 with a technique called the Merlin-Arthur protocol. The training process presents prompts with supporting context partially obscured, teaching the model to verify evidence and return explicit abstentions when documentation is missing. On the AA-Omniscience evaluation, Kolibri-1 refused or withheld ungrounded answers 44% of the time, compared to 15% for the earlier internal Kolibri Origin iteration.
Benchmark Performance and Trade-Offs
In evaluations published by Aleph Alpha, Kolibri-1 demonstrated strong performance on mathematics and German-language reasoning benchmarks compared to other sparse architectures in the 3-billion active parameter category. Kolibri-1 scored 96.9% on AIME 2025 in English and 87.5% on the translated German AIME 2025 benchmark, outpacing baselines like Nemotron 3 Nano.
However, technical disclosures highlight clear trade-offs. Kolibri-1 scored lower in parametric closed-book knowledge tasks, answering 14.8% of Omniscience questions correctly compared to 22.2% achieved by Qwen3.5-35B-A3B. It also trailed competitors on multi-turn tool calling (scoring 39.8 on the Berkeley Function Calling Leaderboard multi-turn eval) and middle-context spans on the RULER benchmark.
Serving the model requires the dedicated aleph-alpha-inference vLLM plugin. Weights and serving scripts remain freely accessible on Hugging Face, while Aleph Alpha retains proprietary rights to its internal training code.
Frequently asked questions
What license is Kolibri-1 released under?
Kolibri-1's model weights and configuration files are released under the open-source Apache 2.0 license on Hugging Face.
What are the hardware requirements to run Kolibri-1?
Kolibri-1 requires approximately 78 GB of VRAM in FP8 precision. The minimum hardware setup is two 80 GB Nvidia A100 or H100 GPUs, or a single Nvidia H200, B200, or B300 accelerator.
How does Kolibri-1's mixture-of-experts design work?
The model has 78.1 billion total parameters across 50 layers. For each token, an MoE router selects 6 out of 384 specialist experts alongside 1 shared expert, activating only 3.46 billion parameters during inference.
What languages does Kolibri-1 support?
Kolibri-1 is specialized specifically for German and English, utilizing a custom 128,000-token UniBPE tokenizer designed to process German compound words efficiently.
Sources
- Aleph-Alpha/Kolibri-1Hugging Face · Oct 2, 2026 · Official
- Kolibri Has Landed: A Sovereign Open-Weight Model — Aleph AlphaAleph Alpha · Oct 3, 2026
- Aleph Alpha Releases Kolibri for Self-Hosted German and English AIsuperpowerdaily.com · Oct 3, 2026
- Aleph Alpha puts Kolibri's full weights on Hugging Face under Apache 2.0Startup Fortune · Oct 3, 2026
- Aleph Alpha Kolibri: How the Sovereign German LLM WorksTejas Kumar · Oct 3, 2026
How this story was made: the newsroom picked it up from Reddit, Hacker News and Google News, gathered the full text of the sources above, and drafted it with AI assistance. Every factual claim was then checked against those sources before publishing (30 claims checked). Illustrations marked as AI-generated are not photographs. Spotted an error? Tell us.
Published October 4, 2026 at 00:05 UTC
