Updated

Kolibri

Aleph Alpha released Kolibri, an English-German open-weight Mixture-of-Experts language model with 78B total parameters, 3B active parameters, and a context length of up to 1 million tokens. Full weights are available under Apache 2.0; the provider describes it as specialized for regulated, mission-critical work.

By No Code MBA · Facts checked

Provider
Aleph Alpha
Announcement date
Status at last check
Generally available
Model family
Kolibri
Preceding model
Kolibri Origin
Capabilities
Text, Coding

What changed

Kolibri is an English-German Mixture-of-Experts model with 78B total and about 3.5B active parameters, up from Kolibri Origin's 30.6B total and 65K context, which was never publicly released. It supports up to 1M tokens of context, trained at 256K, on nearly 24T tokens with a June 18, 2026 knowledge cutoff and four reasoning effort levels. Aleph Alpha reports 96.9 on AIME 2025, 84.3 on GPQA Diamond and 85.9 on LiveCodeBench v6, and a non-hallucination rate of 44.0 versus Origin's 14.8.

Access and availability

Weights are on Hugging Face at Aleph-Alpha/Kolibri-1 under Apache 2.0, served with the aleph-alpha-inference package and its vLLM plugin or container image. Enterprise deployment is available through Aleph Alpha sales.

Official access documentation

What you can build

Evaluate it locally on representative English and German documents, including long-context retrieval and abstention cases, before considering it for regulated workflows.

Limitations

It covers only English and German, and serving beyond 262,144 tokens needs manual configuration, with no published results above 256K. Benchmarks use Aleph Alpha's own harnesses, and Qwen3.6-35B-A3B leads it on some long-context and agent tests. It needs at least one H200 or B200, or two A100 80 GB GPUs, and is not designed for unsupervised operation.

Sources and release history