2026-10-16
Kolibri-1: the German 78B model that fits on one GPU
Aleph Alpha releases an open-weight MoE for German and English, with 3.46 billion active parameters. Here is the architecture and the minimum hardware, from one H200 to two H100s.
Kolibri-1 is Aleph Alpha's open-weight model for German and English assistants. It was released on 3 October 2026 under the Apache 2.0 license and joins the SophyAI catalog with three reference profiles.
Architecture
It is a Mixture of Experts with 78.1 billion parameters, of which 3.46 billion are active per token. It has 50 layers and 384 experts per layer: the router uses 6, plus one shared expert. Attention alternates local windows and global attention in a 4-to-1 ratio. The tokenizer is built around German word structure.
| Feature | Value |
|---|---|
| Total / active parameters | 78.1B / 3.46B |
| Languages | German and English |
| Input / output | text |
| Native context | 262,144 tokens |
| Validated context | up to 1,048,576 tokens |
| Precision | FP8, with an FP8 KV cache |
| License | Apache 2.0 |
Pretraining used about 20 trillion tokens: 62.5% English, 23.9% German and 13.6% code. The knowledge cutoff is 18 June 2026. Reasoning and tool calling are controlled by the template: effort can be low, medium, high or off.
Aleph Alpha is a signatory of the EU GPAI Code of Practice. That does not make a deployment compliant by itself: the system that embeds the model still has to be assessed for its use case.
Memory and hardware
The stated footprint is about 78 GB. The Hugging Face repository occupies 78.9 GB. The published minimum is therefore:
- 2× A100 80 GB or 2× H100 SXM5;
- 1× H200, 1× B200 or 1× B300.
The recommended configuration is 2× H100, 2× H200, 1× B200 or 1× B300. A single 80 GB H100 cannot hold the weights.
Recommended serving stays within 262,144 tokens. Quality has been validated up to 1M, but Aleph Alpha advises against that context when latency and complex tasks matter. Extending it requires --max-model-len 1048576 and an override of max_position_embeddings.
The runtime is not stock vLLM. It needs the aleph-alpha-inference plugin, or the ghcr.io/aleph-alpha/aleph-alpha-inference image, with the kolibri1 parser for reasoning and tool calling.
SophyAI profiles
Three recipes are visible and not yet launchable:
- vLLM FP8 on 1× H200, with a 262k serving context.
- vLLM FP8 on 1× B200, the recommended single Blackwell GPU.
- vLLM FP8 on 2× H100, when H200 and B200 are unavailable.
They are reference: we do not publish SophyAI prices or tok/s, because Aleph Alpha provides no absolute measurement and we have not completed a session. Community GGUFs exist, including a 47 GB Q4, but we do not catalog them until the runtime is verified.