Aleph Alpha releases Kolibri-1, a German-English MoE reasoning model
78B parameters, 3.46B active. Apache 2.0. A sovereignty bet dressed as efficiency.
The facts
- Kolibri-1 has 78.1B total parameters but activates only 3.46B per token across 384 experts per layer (1 shared, 6 routed)
- Licensed Apache 2.0, with a context length of 1,048,576 tokens (recommended limit 262,144 for serving efficiency)
- Trained on 20T tokens pre-training (62.5% English, 23.9% German, 13.6% code), plus 3.44T mid-training and 201B long-context tokens, using 768 NVIDIA B200 GPUs over 21 days for pre-training
- Aleph Alpha reports Kolibri scores 75.5 overall (EN) and 70.8 overall (DE) on its post-training evaluation, ahead of comparably-active-parameter MoE models like GLM-4.7 Flash and Nemotron 3 Nano
- Minimum hardware requirement is 2x A100 80GB, 2x H100 SXM5, 1x H200, 1x B200 or 1x B300; model weights use FP8 quantization
Aleph Alpha has released Kolibri-1, a mixture-of-experts reasoning model focused on German and English, under an Apache 2.0 licence. The model has 78.1 billion total parameters. It activates only 3.46 billion per token, routing across 384 experts per layer, with one shared and six routed experts selected per pass. The design aims to give the capacity of a much larger model at the inference cost of a small one.
The model supports an explicit reasoning mode, with configurable effort levels: low, medium, high, or none. It also supports Hermes-style tool calling. Both are controllable through the chat template at serving time. Context length runs to 1,048,576 tokens. Aleph Alpha recommends capping at 262,144 tokens for serving efficiency and complex tasks. But it says quality and serving efficiency have been validated up to the full million-token window. Most attention layers in the 50-layer transformer use sliding-window attention over nearby text. Only a smaller subset attends across the full context. The company says this keeps long-context inference affordable.
Aleph Alpha says pre-training ran on 20 trillion tokens of filtered bilingual data: roughly 62.5% English, 23.9% German, 13.6% code. Mid-training added 3.44 trillion tokens, and long-context extension a further 201 billion. Pre-training used 768 Nvidia B200 GPUs over 21 days, a total of 392,000 GPU-hours, for 6.4×10²³ FLOPs. The company also built a tokenizer tailored to German word structure. It says this lets German text be processed efficiently without hurting English performance.
On its own post-training evaluation, Aleph Alpha reports Kolibri scoring 75.5 overall on English tasks and 70.8 on German tasks. That puts it ahead of other mixture-of-experts models with similarly sized active-parameter counts, including GLM-4.7 Flash 30B-A3B and Nemotron 3 Nano 30B-A3B. Dense models with far more active parameters per token, such as Qwen3.8 27B, still score higher on some metrics. Minimum deployment hardware is two A100 80GB GPUs, two H100 SXM5s, one H200, one B200 or one B300; FP8 weights have a memory footprint of about 78GB. Aleph Alpha, a signatory of the EU's GPAI Code of Practice, lists a release date of 3 October 2026.
Why it matters
Kolibri-1 is a bet that depth beats breadth. Rather than chasing the broad multilingual coverage of Qwen or Gemma, Aleph Alpha is targeting European enterprises whose workloads run overwhelmingly on German and English. These buyers may value the sovereignty story and the open weights over marginal gains elsewhere. Teams already running small-active-parameter MoE models like GLM-4.7 Flash or Nemotron 3 Nano on comparable hardware should treat Kolibri-1's reported scores and Apache 2.0 terms as grounds for a direct swap-in test, not a wait-and-see.
Written from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.