Kolibri 1
modelYour notes
A German-and-English mixture-of-experts reasoning model built by Aleph Alpha Research GmbH and released on October 3, 2026, the Day of German Reunification. Aleph Alpha signed a definitive business-combination agreement with Cohere in September 2026 — dual headquarters in Berlin and Toronto, Heidelberg kept as a research centre, and the combined company to operate under the Cohere brand — so Kolibri is filed here, though it ships under the Aleph Alpha name and the transaction had not yet closed at release.
78B total parameters with 3.46B active per token: 50 layers, 4:1 sliding-window to grouped-query attention, and 384 experts per layer with one shared and six routed, trained with Muon and what the team calls Exact Quantile Balancing. Native context is 262,144 tokens and quality was validated to 1,048,576, which works because positional encoding is applied only in the sliding-window layers, so the context extends without position scaling; Aleph Alpha still recommends staying at or below 262K for serving efficiency. Weights ship in FP8 (float8_e4m3fn, 128×128 blocks, dynamically quantized activations and an FP8 KV cache) with embeddings, LM head, norms and the MoE router in bfloat16, which puts the footprint near 78 GB — one H200, B200 or B300, or two H100s. Apache 2.0, with an explicit reasoning mode and tool calling, and a June 18, 2026 knowledge cutoff.
Pretraining ran on 20T tokens of a filtered bilingual corpus (about 62.5% English, 23.9% German, 13.6% code), then 3.44T tokens of mid-training and 201B for long-context extension, on 768 NVIDIA B200s across 96 HGX nodes — 21 days and 392,000 GPU-hours for pretraining, 90,000 more for mid-training and 10,000 for long context, about 6.4×1023 FLOPs in total. Aleph Alpha stresses that training happened in Germany and Finland under European and German law with no foreign control, and that it owns the pipeline end to end; it is a signatory of the EU GPAI Code of Practice. Against Kolibri Origin (its 30B-A3B pipeline-validation predecessor), Qwen3.6-35B-A3B, Nemotron 3 Super 120B-A12B and Mistral Small 4 119B-A6B, the launch table reports AIME 2025 96.9 (81.9 / 84.6 / 91.7 / 79.8) and 87.5 in German, AIME 2026 96.0, GPQA Diamond 84.3, LiveCodeBench v6 85.9, HumanEval+ 92.7 and BFCL v4 61.4 — ahead on reasoning and maths, behind Qwen3.6 on BFCL. On the agentic-RAG benchmark Honeypot it beats every model compared, including the much larger Nemotron 3 Super, and on a cleaned MuSiQue it places second to it. The target is sovereign, regulated work: public administration, industrials and aerospace.
Model Details
Benchmark Scores
| Benchmark | Score | Mode |
|---|---|---|
| AIME 2025 | 96.9 | — |
| AIME 2025 (German) | 87.5 | — |
| AIME 2026 | 96.0 | — |
| GPQA Diamond | 84.3 | — |
| LiveCodeBench v6 | 85.9 | — |
| HumanEval+ | 92.7 | — |
| BFCL v4 (overall) | 61.4 | — |
Variants
| Name | Parameters | Notes |
|---|---|---|
| Kolibri-1 (FP8) | — | float8_e4m3fn weights in 128x128 blocks with an FP8 KV cache; ~78 GB footprint |
| Kolibri-1-BF16 | — | bfloat16 release the FP8 build is quantized from |
| Kolibri Origin | — | 30B total / 3B active, 65K context; built first to validate the training pipeline, released about three months earlier |