Kolibri model: 78B MoE, 3.46B active, Apache 2.0 weights for German and English

Aleph Alpha positions Kolibri for regulated work in public administration, industrials and aerospace. The model was built in Germany and trained on infrastructure in Germany and Finland. For teams whose workload is mostly German and must stay on their own hardware, there is now an Apache 2.0 option with a bilingual German/English tokenizer and a published account of how the training data was curated. The price is holding 78B parameters in memory to get 3B-class compute per token.

The sources disagree on training tokens. The launch post says 20T pre-training tokens, distilled from over 200T tokens of raw data. The summary table in the tej.as write-up says about 24 trillion training tokens. The predecessor, Kolibri Origin, was a 30.6B-total, 3.27B-active model with a 65,536-token longest trained length and 7.51T pre-training tokens. It was never publicly released, and its pre-training finished on 11 June, three months before Kolibri's on 11 September.

Tokenizers trained mostly on English text chop up German compound words. GPT-5's o200k_base splits Bundesverfassungsgericht into 6 tokens, while Kolibri's tokenizer uses 2. Per the tej.as summary, UniBPE keeps the bottom-up merging of byte-pair encoding but scores each merge with the Unigram objective. Aleph Alpha's technical report says this needs 11.2% fewer tokens for German text than GPT-5's tokenizer, the best of the 9 others it measured.

The tej.as author ran an independent check on the full German Basic Law. Kolibri used 35,190 tokens, against 41,482 for o200k_base, 42,907 for Qwen3.5 35B-A3B, 43,478 for Mistral Small 4 and 43,850 for Gemma 4. That is about 15% fewer than GPT-5's tokenizer. On the official English translation it was effectively tied with o200k_base (39,875 vs 39,737). Fewer tokens means fewer decode steps for German output and more German per context window, but the test covers one legal document, and one Hacker News commenter asked whether the tokenizer is better in practice.

Aleph Alpha compares Kolibri with Qwen3.6-35B-A3B, Nemotron 3 Super 120B-A12B and Mistral Small 4 119B-A6B. Kolibri posts AIME 2025 96.9 (87.5 in German), GPQA diamond 84.3 (81.3 in German), LiveCodeBench v6 85.9 and τ³-bench banking 38.1, where the next best is Nemotron 3 Super at 15.5. Against Nemotron 3 Super, which has about four times its active parameters, Kolibri scores at or above it on every listed benchmark except HumanEval+ (92.7 vs 94.7).

Kolibri is behind Qwen3.6-35B-A3B on LongBench Pro (64.5 vs 70.8), AA-LCR (68.3 vs 69.7), BFCL v4 (61.4 vs 67.2), τ²-bench telecom (94.7 vs 99.1) and the AA-Omniscience Index (-32.8 vs -15.3). Aleph Alpha also reports internal customer-proxy suites. On the German public sector suite, scores rose from 0.54 to 0.75 across successive post-training runs. Those suites are internal, and every number in this section is vendor-run.

The model card is explicit that the full model must be held in memory even though only part of it is active at any time. That comes to about 78 GB of weights in FP8. Aleph Alpha's serving argument is concurrency. On two H100s, a 123B variant it tested handled only 3 long-context 256k-token queries, while the 78B design handles 18 concurrent requests and decodes 28% faster. The four reasoning levels add a second lever, trading cost and latency against answer quality per request.

Neither the launch post nor the tej.as write-up, in the portions reviewed here, mentions official quantized releases, quality loss under quantization, or support in specific inference engines. That gap matters for anyone sizing hardware below the FP8 footprint.

One commenter on Hacker News questioned the intellectual-property-safety claim, arguing that a competitive model cannot be built without unlicensed training data. Other commenters disputed that. The sources partly speak to provenance. According to tej.as, the model card says English web text was rephrased with Gemma 4, German text with Mistral-NeMo, and Qwen3-32B labeled data for the quality filters, with the data then filtered for political bias. As that write-up puts it, sovereign does not mean nothing from outside Europe went in.

Another commenter judged Kolibri to perform about like Qwen3.6 35B A3B on the benchmarks. They said you would have to be a bit desperate to use it for coding, but suggested it would work for information extraction or classification in a regulated industry. They guessed it would run in roughly 96 GB of RAM with a decent quant, a figure no source verifies. A third commenter said Aleph Alpha recommended high-end setups without considering 4- or 8-bit quantization.

The sharpest practical complaint came from a user of a third-party hosted demo. They asked the model how to run itself on llama.cpp with less RAM than stated. With extended thinking on, they got little beyond advice to check the Hugging Face page. That is the flip side of the abstention training Aleph Alpha built with its Merlin-Arthur protocol, which teaches the model to say it does not know when the answer is not in the context. Neither source reports how often that abstention costs a correct answer.

Kolibri makes sense when three conditions hold. Most of your input and output is German. The deployment has to stay on-premise under an Apache 2.0 licence, with a model built in Germany and trained in Germany and Finland. And you can hold about 78 GB of FP8 weights plus whatever your context lengths require. In that setting the tokenizer savings, the German AIME and GPQA scores, the banking and airline agent results, and the grounding-oriented abstention behaviour all favour it.

Qwen3.6-35B-A3B is the better default for long-document retrieval, function calling as measured by BFCL v4, and tighter memory budgets. Its name indicates about 35B total parameters, against Kolibri's 78.1B. Both comparisons rest on Aleph Alpha's own benchmark runs.

Neither source reports independent benchmark reproductions, and the sector suites are internal. No source reviewed here mentions official quantized weights, quality-loss figures or support in particular inference engines. Quality between the 262,144-token trained length and the 1,048,576-token tested length is not documented in the reviewed material, and neither source gives a false-abstention rate for grounded question answering. The 20T versus 24T training-token figure is unresolved, so teams evaluating Kolibri should run their own German workloads at the quantization and context length they intend to serve.