Kolibri Models by Aleph Alpha
Last refreshed 2026-10-05. Next refresh: weekly.
Details
Capabilities
About
Kolibri is Aleph Alpha's open-weight family of English–German mixture-of-experts reasoning models, trained from scratch in Germany for sovereign, regulated deployments (public administration, industry, aerospace). The first public model, Kolibri 1 (released 3 October 2026), has 78B total and 3.46B active parameters, a native 262K-token context validated to 1M tokens, controllable reasoning, and tool calling, with Apache 2.0 weights on Hugging Face. It is a separate lineage from Aleph Alpha's proprietary Luminous line and the Pharia-1 releases, built in its Model Factory pipeline; the internal predecessor Kolibri Origin was not released.
Decision facts
- Best fit
- agentcodingreasoning
- Capability starting point
- Kolibri 1 with 1.05m context and reasoning, JSON / Tool use, and structured outputs
- Lowest tracked input
- Not tracked
- Closest related family
- Aleph Alpha Luminous
Current Variants
Use-when guidance is based on each model's tracked capabilities, context window, release date, and replacement status.
Use when the workload needs 1.05m context, reasoning, and JSON / Tool use.
| Model | Use when | Released | Signals | Status |
|---|---|---|---|---|
| Kolibri 1 | Use when the workload needs 1.05m context, reasoning, and JSON / Tool use. | 2026-10 | 1.05m contextreasoningJSON / Tool use | Current |
Release Timeline
1 release groupSpecifications(1 models)
| Model | Released | Context | Parameters | Reasoning | JSON / Tool use | Structured Outputs |
|---|---|---|---|---|---|---|
| Kolibri 1 | 2026-10 | 1.05m | 78B (3.46B activated) | Yes | Yes | Yes |
