Aleph Alpha releases Kolibri-1, an open-weight German-English reasoning model
The short version
Aleph Alpha released Kolibri-1 on October 3. The Apache 2.0 MoE model has about 78B total parameters and about 3.46B active per token, supports tool calling, and offers a maximum context of 1,048,576 tokens with a lower recommended operating length.
What changed
Weights and a model card are now public for self-hosted evaluation.
What it means for you
The model offers a German-English deployment option, but its active parameter count does not remove the memory needed for full weights.
Kolibri-1 targets German and English reasoning and tool-assisted tasks. Its model card specifies 78,103,074,560 total parameters and 3,457,573,120 active per token; the launch post rounds the active figure to 3B.
The maximum context is 1,048,576 tokens, while the card recommends at most 262,144 for serving efficiency and complex tasks. Native long-context training is 262,144 tokens. Accepted input length should not be confused with demonstrated accuracy across an entire document.
The main checkpoint is FP8 and the card estimates approximately 78 GB for weights. Minimum examples include two 80 GB A100 GPUs, two H100 SXM5 GPUs, or a single H200, B200 or B300. Serving overhead and caches require additional capacity. The separate BF16 checkpoint has different requirements.
The next useful step is a small German-English evaluation with a conservative context cap. Check citations, tool parameters and measured memory before deployment. The catalog entry links to the primary model card. Event time uses the October 3 date on the official launch post, without inventing a time of day.
Fact check
- VerifiedAleph Alpha released Kolibri-1 on October 3, 2026 for German and English reasoning and tool use.Evidence
- VerifiedThe model has 78,103,074,560 total parameters and 3,457,573,120 active per token, and Apache 2.0 weights.Evidence
- VerifiedMaximum context is 1,048,576 tokens; the recommended cap and native long-context training length are 262,144.Evidence
- VerifiedThe FP8 card estimates 78 GB weight memory and lists two 80 GB A100s, two H100 SXM5s or one H200/B200/B300 as minimum examples; BF16 is separate.Evidence
Coverage timeline
- Primary sourceAleph AlphaKolibri Has Landed: A Sovereign Open-Weight Model
- Primary sourceAleph AlphaKolibri-1 model card
