Kolibri 1: Inside Aleph Alpha's Open-Weight LLM for German and English

On 3 October 2026, Aleph Alpha published the weights of Kolibri 1 on Hugging Face under the Apache 2.0 license: a mixture-of-experts reasoning model that, by the claim that matters in Germany, is the country's first open-weight LLM of this class. Christoph Magnussen's video 'KI aus Deutschland! Was steckt hinter Kolibri von Aleph Alpha?' works through the model the same way this report does, against the primary source, the model card, rather than against demo impressions. The headline numbers are 78 billion total parameters with 3.46 billion active per token, a validated context window of 1,048,576 tokens, and a 20-trillion-token pre-training corpus in which 23.9 percent is German. What makes the story interesting is the combination: an open license, a deliberate two-language scope, and a deployment path that starts at rented or owned GPUs, not at a 20-euro chat subscription.

The card describes a 50-layer MoE transformer: 384 experts per layer, one shared and six routed, trained with the Muon optimizer and Exact Quantile Balancing, mixing sliding-window and grouped-query attention at a 4:1 ratio. Total parameters come to 78,103,074,560, active per token to 3,457,573,120. Weights ship in float8_e4m3fn with 128x128-quantised blocks, while embeddings, the LM head, norms and the router stay in bfloat16, and the KV cache runs in FP8; the full footprint is roughly 78 GB, and that number decides where the model can live. Context is validated to 1,048,576 tokens, with 262,144 recommended for serving and complex tasks, and the sliding-window design extends the context without positional rescaling. Two modes are named explicitly on the card: a reasoning mode and tool calling. The knowledge cutoff is 18 June 2026 for both languages. One detail with practical weight: the tokenizer is built for German word structure, so German input is processed efficiently without giving English worse treatment.

The compute bill reads like a status report. Pre-training ran on 768 NVIDIA B200 units across 96 HGX nodes with 8 GPUs each, took 21 days, 511 hours, about 392,000 GPU-hours on 20T tokens, split 62.5 percent English, 23.9 percent German and 13.6 percent code, mixed with synthetic rephrasings, translations and curated high-quality sources. Mid-training added 3.44 trillion tokens over five days and 90,000 GPU-hours, and the long-context phase added 201 billion tokens in 13 hours and 10,000 GPU-hours. The tech report totals it at 6.4e23 FLOPS with an energy estimate around 950 MWh including datacentre overhead via PUE, excluding SFT and RL. On the video's side, Magnussen stresses the location claim, training on European hardware, and confirms the language share: about a quarter German, consistent with the card's 23.9 percent.

The card lists the house results: GPQA Diamond 84.3, AIME 2026 96, SWE-bench Verified 66.4, Terminal-Bench 2.1 27.7 and Humanity's Last Exam 21.5, plus an MMLU-Pro 80 with a note attached. Solid numbers for the size class, but they are the vendor's own measurements; no independent replication has surfaced yet. A throughput test in the NVIDIA developer forum measured single-stream generation around 47 to 49 tokens per second on B200-class hardware, which matters for serving plans, not for model league tables. The video deliberately skips benchmark chest-beating and frames Kolibri as a model for critical applications. The thinnest part of the evidence base is, ironically, the German language: independent evaluations of German office or bureaucratic language do not exist yet, exactly where the sovereignty argument should pay off.

There is no play button. The official path is the aleph-alpha-inference package with its Kolibri vLLM plugin, or the container image ghcr.io/aleph-alpha/aleph-alpha-inference; a typical start reads vllm serve Aleph-Alpha/Kolibri-1 with flags for an FP8 KV cache, a 262,144-token model length, the kolibri1 reasoning parser and the matching tool-call parser. On the hardware side the card sets the floor at two A100 80 GB, two H100 SXM5, or a single H200, B200 or B300, and recommends the tier above. Magnussen puts the entry ticket at 30,000 to 40,000 euros, and his own MacBook Air with 32 GB RAM fails the LM Studio check with 'likely too large'. A community-quantised, roughly Mac-sized variant exists on Hugging Face, but the realistic ceiling on Apple silicon is a MacBook Pro with 128 GB of RAM or a Mac Studio. The unofficial routes: a company near Munich exposes the model at tesseracted.com, with request limits and no automatic GDPR guarantee, or you rent a Hugging Face Space and operate the model there. Both work, says the video, and both are operations work.

The intended-use section is sober: document processing, question answering over an organisation's own material, internal knowledge and research tools, structured extraction and retrieval-augmented generation, agentic tool calling where the calling system validates results, and systems in which a person reviews output before it is acted upon. The model belongs on the advisory side of decision support and is explicitly not meant to be the deciding component. The video's scenario list runs from meeting transcription and CRM notes to the sensitive domains, defence, healthcare and cyber defence. The often-missed point: Apache 2.0 permits commercial use, modification and redistribution with no user or revenue thresholds, and open weights make the model inspectable. None of that makes an operation GDPR-compliant by itself, as the video says in plain terms; the environment, access rules and audit processes still have to be engineered.

Aleph Alpha, based in Heidelberg, was founded in 2019 and has been through a real pivot: long marketed as Germany's answer to OpenAI, the company moved its focus, per the video's telling around 2024 or 2025, away from building frontier models and toward running sovereign AI systems for government and industry. A few weeks before the video, Aleph Alpha announced a merger with the Canadian model builder Cohere, says Magnussen; it has not closed yet, and the valuation history had its ups and downs. The model card documents one institutional fact: Aleph Alpha is a signatory of the EU GPAI Code of Practice. Magnussen's strategic point is a double negative: if Europe simply declares the LLM race over, it does not lose the race, it forfeits participation, and the capability that matters builds up by running local models. His side notes frame Kolibri against the field: Chinese models often reason internally in English, and models like Astra or Fable, per the video, reveal much less of their reasoning than an explicit reasoning mode does.

For whom is this relevant? Operators, agencies and system integrators who want critical workflows on their own infrastructure and are ready to treat inference as a project, with GPU budget, privacy engineering and their own evaluations. Consumers get nothing to subscribe to, and the video keeps the honest wording: 'could', not 'can'. There is a market for locally operated models, but it needs different people than the 20-euro chatbot audience, infrastructure engineers, compliance officers, operations. Germany's claim on Kolibri 1 is therefore less a victory lap than a starting point: the weights are public, the license is permissive, the German share of training is real, and the value will be decided in deployment, not in benchmarks. Three things remain open: the Cohere merger, an officially hosted inference offering, and independent German-language evaluations. Until then, the model card and the tech report are the contract.

Video: https://www.youtube.com/watch?v=SOcdaG5KZbs
Model card: https://huggingface.co/Aleph-Alpha/Kolibri-1
Tech report: https://aleph-alpha.com/downloads/tech-report.pdf
Aleph Alpha blog: https://aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model/

YouTube video 'KI aus Deutschland! Was steckt hinter Kolibri von Aleph Alpha?' by Christoph Magnussen (channel Christoph Magnussen, Blackboat), accessed 12 Oct 2026.
Hugging Face, model card Aleph-Alpha/Kolibri-1, accessed 12 Oct 2026.
Aleph Alpha, 'Kolibri Has Landed: A Sovereign Open-Weight Model', blog, accessed 12 Oct 2026.
Aleph Alpha, tech report (PDF), accessed 12 Oct 2026.
NVIDIA developer forum, Kolibri-1 benchmark thread on B200, accessed 12 Oct 2026.