Aleph Alpha Kolibri: How the sovereign German LLM works

What it claims

Kolibri-1 (Aleph-Alpha/Kolibri-1 on Hugging Face) is a 78.1B-parameter MoE with 3.46B active per token. It covers German and English only, ships under Apache 2.0, and was released 3 October 2026. Aleph Alpha says it beats every open MoE of its active-parameter class in both languages. The post's author, who is not affiliated, independently checked one claim: the tokenizer efficiency. The author did not run the model itself.

How it works

• MoE layout: 50 layers, each with 384 routed experts plus 1 shared expert. The router sends each token to the shared expert and 6 routed experts. All 78B weights must stay in memory.
• Attention: 40 of 50 layers use sliding-window attention over the previous 512 tokens. Every 5th layer attends to the full context.
• Only the sliding-window layers carry rotary position embeddings. Because of this, context extends past the trained 262,144 tokens with no position-scaling tricks. It is validated to 1,048,576.
• Tokenizer (UniBPE): 128,000-token vocabulary. Merges are built bottom-up as in BPE but scored with the Unigram objective.
• German reasoning: about 800,000 generated German reasoning traces.
• Aleph Alpha found that a small amount of such data hurts: German math fell from 70.2 to 48.3 because reasoning looped without finishing.
• It recovered only to 67.3, and only with much more data.
• Merlin-Arthur abstention training: Arthur (the model) gets a question with parts of the source document hidden.
• Merlin hides parts so that the answer stays findable, and Arthur is trained to answer.
• Morgana hides the evidence the answer depends on, and Arthur is trained to say he doesn't know.
• Arthur can't tell which one he's facing, so he must check whether the visible evidence supports an answer.

flowchart LR
  Q[Question + document] --> M{Who masks?}
  M -->|Merlin: keeps evidence| A1[Train: answer]
  M -->|Morgana: removes evidence| A2[Train: say 'I don't know']
  A1 --> AR[Arthur can't tell which]
  A2 --> AR

Evidence

All results come from Aleph Alpha's own harness, using each vendor's recommended sampling, except where noted.

• AIME in English (2025 and 2026): Kolibri beats every MoE in the comparison, including 12B-active ones. Only dense Qwen3.8 27B scores higher.
• Company-document tests: 5 internal tests covering semiconductors, German public sector, aerospace, an automotive supplier and industrial drives. All use unseen documents. The post gives no numbers for them.
• Author's independent tokenizer test:
• Corpus: the full German Basic Law (185 KB) and its official English translation.
• Kolibri needed 15% fewer tokens than o200k_base (GPT-5) on the German text and tied on English.
• Aleph Alpha's own figure for German is 11.2%.
• Example: "Bundesverfassungsgericht" is 2 tokens in Kolibri versus 6 in o200k_base.

How to use it

Hardware: about 78 GB of GPU memory. That means 2× A100/H100 80 GB, or a single H200, B200 or B300.

Install and serve:

pip install 'aleph-alpha-inference>=1'
vllm serve Aleph-Alpha/Kolibri-1 --kv-cache-dtype fp8 \
  --reasoning-parser kolibri1 \
  --tool-call-parser kolibri1 \
  --enable-auto-tool-choice

This gives an OpenAI-compatible endpoint. Reasoning effort (none, low, medium, high) is set through the chat template:

extra_body={"chat_template_kwargs": {"reasoning_effort": "high", "enable_thinking": True}}

Recommended settings:

• Sampling: temperature=1.0, top_p=0.97, top_k=128.
• Keep context at or below 262,144 tokens for latency-sensitive use.
• For the full 1M context:

vllm serve Aleph-Alpha/Kolibri-1 --kv-cache-dtype fp8 \
  --max-model-len 1048576 \
  --hf-overrides '{"max_position_embeddings": 1048576}'

Limits

Weak spots (Aleph Alpha's own numbers):

• Closed-book knowledge: last of 12 models on RGB with no documents (51.0). Only 14.8% of Omniscience questions answered correctly, versus 22.2% for Qwen3.5 35B-A3B. It works with documents in the prompt, not as a trivia oracle.
• Multi-turn tool calling: BFCL multi-turn 39.8, versus 58.2 for GLM-4.7 Flash and 54.0 for Qwen3.5 35B-A3B.
• Coding agents: Terminal-Bench 2.1 27.7 versus 39.7 for Qwen3.5 35B-A3B. SWE-bench Verified 66.4 versus 73.8 for Qwen3.6 35B-A3B.
• Mid-length context: RULER at 128k is 67.9 versus 89.9 for Qwen3.5 base. Kolibri only leads at the very long end.

Deployment constraints:

• It needs Aleph Alpha's vLLM plugin, which pins a single vLLM version (0.29 today).
• No hosted provider served it on launch day.
• Apache 2.0 covers the weights and config files only. Training code and methods are not released.

Data provenance: "Sovereign" does not mean non-European inputs were excluded. Gemma 4 rephrased the English web text, Mistral-NeMo rephrased the German, and Qwen3-32B labeled data for the quality filters.