Germany's Free Open Model: What Kolibri-1 Actually Offers

Germany has a new free-model claim that is turning heads. The host says Kolibri beats well-known open models such as Qwen, Nemotron and Mistral, and walks through the setup step by step on video. The name sounds like Calibri in the narration, but the official records spell it Kolibri-1 . This article lays out what the Kolibri-1 model actually offers, which scores are verified, and the hardware cost you must know before downloading. The narrative follows the video's order and tests every claim against independent records.

The company behind the model is Heidelberg-based Aleph-Alpha. On 3 October, Germany's Unity Day, it published the full weights on Hugging Face under the permissive Apache 2.0 license. The license covers the weights and configuration files; according to the company, the model focuses on text generation and runs in German and English only. The launch date looks deliberate, since Aleph-Alpha frames the release as sovereign European infrastructure. This account is drawn from Aleph-Alpha and preserves the announcement date and license terms exactly.

MoE layout and the 1M-token window

On the technical side, the model is described as a mixture-of-experts Transformer. Of 78B total parameters, only 3.46B active parameters engage per token; the configuration record notes 6 out of 384 experts selected per token. Weights sit in FP8 precision with a total footprint around 78 GB. The context window reaches 1M tokens, though 262K tokens is recommended as the efficient serving ceiling. The model ships with an explicit reasoning mode plus tool-calling support. These figures are drawn from HuggingFace and match the parameter and context counts on the model card exactly.

Training scale is reported in striking numbers. Pre-training started on 20T tokens, followed by 3.44T tokens of mid-training and 201B tokens of long-context extension. The German share reads differently across official texts, between 20% and 23%; the current model card gives 23%. The knowledge cutoff is recorded as June 2026, so offline use needs tool support for anything newer. Training reportedly ran on 768 B200 accelerators in data centers in Germany and Finland. This training breakdown is drawn from DataCamp and stays consistent with the company's token and accelerator counts.

Scores and the vendor-measurement caveat

On the scoreboard, the model leads Qwen, Mistral and Nemotron variants on the German average. A similar lead is claimed for agent and tool-calling measures, and the host opens that invitation especially to local-model trials with helpers such as Hermes Agent. Yet the comparison values come from the company's own harnesses, with no independent verification published so far. The English overall score even trails the dense Qwen rival, as the record shows. This cautious reading is drawn from Traictory and keeps the vendor's own measurement clearly apart from verified values.

The critical point before downloading is machine power. The model weights occupy about 78 GB ; the model card recommends at least one H200 or two 80 GB A100 cards plus a vLLM plugin. The host demonstrates setup with LM Studio, but not every machine carries that load; on-card fast storage and driver fit must be checked first. Tool-calling and abstention behavior should also be measured on your own data. This setup warning is drawn from TestMuai and gives the pre-self-hosting checklist in concrete items.

A practical result for firms and public bodies

The final frame is wider. Aleph-Alpha presents the Kolibri model as a sovereign AI example that companies and public bodies can run on their own infrastructure. The promise of EU-compliant local operation rests on reducing dependence on US providers. Even so, the picture has missing pieces: some scores were measured only on company harnesses, and non-text output is unsupported. This wider frame is drawn from Heise and recommends guarded optimism until independent measurements appear.