Kolibri 1: German and English Open-Weight MoE Model
Use Kolibri 1 for self-hosted bilingual text and tool-calling workflows. Compare its sparse computation with the full memory and serving requirements.
Official-source review · Last checked October 4, 2026 · No independent product benchmark
What Kolibri 1 does
Kolibri 1 is Aleph Alpha's German/English MoE model. It accepts text and returns generated text, with reasoning and tool-call support for integrated applications. Consider it for bilingual assistants, coding and document workflows when you have deployment expertise. Official model card.
Compute is not the same as model memory
The card lists 78B total parameters and about 3.46B active per token. Sparse activation reduces computation per token; it does not mean loading only 3.46B weights.
Official specifications: approximately 78 GB of FP8 weights, with selected components in BF16 and evaluation using an FP8 KV cache. One H200 is a listed minimum option; recommended configurations include two H200s. The card recommends contexts of at most 262,144 tokens for efficiency and complex tasks, despite reporting validation up to 1,048,576. Hardware and precision table.
A minimum hardware listing is not a throughput guarantee. Budget separately for runtime overhead, KV cache, concurrent requests and your target context. We have not verified maximum context or production concurrency on a single H200.
Access, license and operating cost
The repository lists Apache 2.0 weights and provides configuration, tokenizer and serving instructions using the Aleph Alpha vLLM plugin. Public weights do not establish that every training dataset or training-code component is released. Review the published files before building a reproducibility claim.
Self-hosting still costs hardware, energy and operations. No hosted per-token price is established here. Choose it for control over a model integration; if you need a ready-to-use assistant without GPU administration, compare Gemini or the chatbot category.
Evaluate before deploying
Aleph Alpha reports benchmark results in its card; AI Tool Finder has not reproduced them. Test the German and English documents your users actually handle, including failures, tool arguments and unsupported claims. Require review before consequential actions.
European origin is not a compliance certificate for your application. Data handling, deployment location, access controls and the intended use need their own assessment. Open weights also do not remove hallucination or bias risks.
Agent frameworks and integration choices · Coding & Development tools