Kolibri 1
Kolibri 1 is a German-and-English mixture-of-experts reasoning model. It has 78.1B total parameters and 3.46B active per token. The downloadable release includes FP8 and BF16 packages.
Model design
The 50-layer transformer combines sliding-window and grouped-query attention in a 4:1 ratio. Each MoE layer has 384 experts with one shared and six routed experts. Pretraining used 20 trillion tokens, followed by mid-training and context extension. The model supports native tool calls.Kolibri 1 model card
Practical use
Practical guidance. Use Kolibri when German wording and institution-specific documents are central to the task. Start within the recommended serving context instead of filling the maximum window. Measure complete tool workflows and unsupported-answer behavior. Active parameter count describes computation, not the memory needed to store all experts.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | Downloadable |
| Model identifiers | Aleph-Alpha/Kolibri-1Aleph-Alpha/Kolibri-1-BF16 |
| Context window | 1,048,576 tokens Maximum model context; the card recommends no more than 262,144 tokens for efficiency and complex tasks. Kolibri 1 model card |
| License | Apache 2.0Kolibri 1 model card |
| Reasoning controls | Low, medium, high or disabled.Kolibri 1 model card |
Related models
- Pharia 1 LLM 7B Control Aligned
Aleph Alpha · Pharia 1
Safety-aligned seven-billion-parameter Pharia 1 model designed for chat applications.
Launch date not confirmedDownloadable · Provider product