Aleph Alpha releases Kolibri, a 78B-parameter (3.5B active) open-weight English–German ‘sovereign’ MoE model under Apache 2.0
Confirmed
The takeaway
On Oct 3, 2026 (German Unity Day) Aleph Alpha released Kolibri-1, an English–German mixture-of-experts model with 78B total and ~3.46B active parameters, a 1M-token context and a German-optimised tokenizer, as open weights under Apache 2.0.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 3 October 2026
Your AI and this story
- GPT-6.2 Astra125 days after its cutoff
- Claude Opus 5.595 days after its cutoff
- Gemini 4 Hydrogen186 days after its cutoff
- Grok 4.7125 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 95 days before it.
Key facts
- Architecture: MoE, 78B total / ~3.46B active parameters, 384 experts per layer with 6 routed per token, 50 layers; sliding-window (512) attention with full attention every 5th layer
- Context: up to 1,048,576 tokens (262,144 recommended; longest trained context 262,144); knowledge cutoff 18 June 2026 (EN/DE)
- Training: ~24T tokens (≈62% English, 14% code, 21.3% / ~4.3T German), curated from >200T raw tokens; 768 Nvidia B200 GPUs in Germany and Finland
- Tokenizer: bilingual EN–DE, 128k vocabulary; Aleph Alpha claims 11.2% fewer German tokens than GPT-5’s tokenizer (an independent test on the Basic Law found ~15%)
- Aleph Alpha’s benchmarks: AIME 2025 96.9 (EN) / 87.5 (DE); GPQA Diamond 84.3 (EN) / 81.3 (DE); HumanEval+ 92.7; compared against Qwen3.6-35B, Nemotron 3 Super 120B, Mistral Small 4, Gemma and GLM models
- Reasoning mode with configurable effort, tool calling, abstention training (Merlin–Arthur ‘M/A grounding score’ 0.23)
- Weaknesses reported by reviewers: closed-book knowledge, long multi-turn tool use and agentic coding trail Qwen models
- Weights: huggingface.co/Aleph-Alpha/Kolibri-1 (Apache 2.0); served via Aleph Alpha’s aleph-alpha-inference package on vLLM; needs ~2×80GB GPUs
What happened
Aleph Alpha published Kolibri-1 on German Unity Day, Oct 3, 2026, as an Apache-2.0 open-weight model on Hugging Face. It is a sparse MoE (78B total, ~3.5B active) built for German and English enterprise and public-sector use. Aleph Alpha says sovereignty “combines two dimensions: how we built the model, and how it transfers to our customers”: the model was trained only on infrastructure in Germany and Finland, and customers can run the weights themselves. A tech report is linked from the blog post.
Aleph Alpha’s own benchmark tables put Kolibri ahead of open models of similar active size on maths (AIME 2025 96.9% in English) and GPQA Diamond (84.3%). Independent write-ups (Tejas Kumar) say it is weaker at closed-book recall, long multi-turn tool use and coding agents. All benchmark numbers are Aleph Alpha’s own.
Why it matters
Kolibri is a notable European open-weight release, from a lab that agreed in Sept 2026 to combine with Cohere. A model with 3.5B active parameters, 1M context and an Apache licence gives German-speaking governments and companies a capable model they can run themselves.
Sources
6 sources from 5 sites. Numbers match the chips in the text.
6 sources: 2 primary, 1 press, 3 reactions
Primary
- Aleph Alpha: Kolibri has landed — a sovereign open-weight modelaleph-alpha.com, official
- Hugging Face: Aleph-Alpha/Kolibri-1 model cardhuggingface.co, code
Press
- Crypto Briefing: Aleph Alpha releases Kolibri, a 78B-parameter open-weight AI model built in Europecryptobriefing.com, press
Reactions
- Tejas Kumar: Aleph Alpha Kolibri — how the sovereign German LLM workstej.as, discussion
- Hacker News discussion (official post)news.ycombinator.com, discussion
- Hacker News discussion (Tejas Kumar explainer)news.ycombinator.com, discussion
Changes
- Filed (two threads)