Aleph Alpha releases Kolibri, an open-weight German and English MoE model

Aleph Alpha released Kolibri on October 3, 2026, an open-weight language model trained for English and German. The German company put the weights on Hugging Face under the Apache 2.0 license and published a long technical post plus a tech report. Kolibri is a mixture-of-experts (MoE) model with 78.1B total parameters, of which 3.46B are active for each token.

What is new in Kolibri

Kolibri follows Kolibri Origin, a smaller model Aleph Alpha built to validate its training pipeline and never released. The announcement post (opens in a new tab) compares the two directly:

• Total parameters: 30.6B for Origin, 78.1B for Kolibri.
• Active parameters per token: 3.27B and 3.46B.
• Pre-training tokens: 7.51T and 20T.
• Longest trained length: 65,536 and 262,144 tokens.
• Experts: 128 total with 8 active, against 384 total with 6 active.
• Reasoning: one mode in Origin, four effort levels (none, low, medium, high) in Kolibri.

The headline context figure is 1M tokens. The Kolibri model card (opens in a new tab) is more precise: the maximum is 1,048,576 tokens, and Aleph Alpha recommends 262,144 tokens or fewer for serving efficiency and complex tasks. Only 10 of the 50 layers attend over the full context. The other 40 use a 512-token sliding window.

German is the second focus. The post says 21.3% of pre-training tokens were German, roughly 4.3T over the 20T run, with translation kept to 6% overall. A new tokenizer method, which Aleph Alpha calls UniBPE, gives a 128,000-token vocabulary. On German web text the company reports 4.90 bytes per token for Kolibri and 4.35 for GPT-5's tokenizer.

Aleph Alpha also trained Kolibri to say "I don't know" when the supplied context lacks an answer, using abstention data and its own Merlin-Arthur procedure. On the public AA-Omniscience set, Kolibri's non-hallucination rate is 44.0, up from 14.8 for Origin.

Benchmarks, as Aleph Alpha reports them

Every score in the post comes from Aleph Alpha's own eval-framework (opens in a new tab), with each model at its highest reasoning effort. Kolibri scores 96.9 on AIME 2025, 84.3 on GPQA Diamond, 85.9 on LiveCodeBench v6 and 66.4 on SWE-Bench Verified. Its overall average is 75.5 in English and 70.8 in German.

The same table shows where Kolibri trails. Qwen3.8 27B, a dense model, averages 80.2 in English and 79.9 in German. Qwen3.6-35B-A3B scores higher on Tau2-Bench Telecom (99.1 against 94.7) and on BFCL v4 overall (67.2 against 61.4). The post cites no independent replication.

Who Kolibri is for

Aleph Alpha names public administration, industrials and aerospace as target sectors. The pitch is on-premise deployment without sending internal data to outside inference services, from a model built in Germany and trained on infrastructure in Germany and Finland. The model card describes intended use as assistants and agent workflows where a person reviews output before anyone acts on it. In decision support, the card places Kolibri on the advisory side.

The small active count keeps compute per token low, but the whole model still has to fit in memory. The card puts the footprint at about 78 GB with FP8 weights. Listed minimum hardware is 2× A100 80 GB, 2× H100 SXM5, 1× H200, 1× B200 or 1× B300. That is a server setup, a different class from the laptop-sized models in our guide to running a local LLM with Ollama.

How to try Kolibri

Kolibri runs on vLLM with a plugin from the aleph-alpha-inference repository (opens in a new tab). Installing the package also installs the vLLM version it supports, which the README gives as vLLM 0.29. A container image is published at ghcr.io/aleph-alpha/aleph-alpha-inference. The commands from the post:

pip install "aleph-alpha-inference>=1.0"
vllm serve Aleph-Alpha/Kolibri-1 --kv-cache-dtype fp8 \
  --reasoning-parser kolibri1 \
  --tool-call-parser kolibri1 \
  --enable-auto-tool-choice

For contexts beyond 262,144 tokens, the post adds --max-model-len 1048576 plus a max_position_embeddings override. Recommended sampling is temperature 1.0, top_p 0.97 and top_k 128. Thinking is on by default; the README shows how to switch it off per request with reasoning_effort: "none" or enable_thinking: false. A BF16 checkpoint is at Aleph-Alpha/Kolibri-1-BF16. If vLLM is new to you, see serving an open model with vLLM behind an OpenAI-compatible API.

Limits and open questions

The Apache 2.0 grant is narrower than some readers will assume. The model card says it covers only the weights and configuration files in the Hugging Face repository. It excludes the underlying code, architecture, parameter settings and training methods, and Aleph Alpha keeps all rights to those.

Pricing is not part of the release. Neither page lists a hosted API or a price, and the post sends enterprise customers to the sales team for deployment and specialization.

The training data is described in detail but not linked as a download on the pages reviewed here. The two primary pages also disagree on the German share: 21.3% in the post, about 23.9% in the model card's training data summary.

The sector results, such as aerospace rising from 0.14 to 0.59, come from internal customer-proxy suites that are not public, so outsiders cannot check them. The card gives a knowledge cutoff of June 18, 2026 for both languages.