Aleph Alpha releases and open-sources the English-German bilingual model Kolibri
BIBIBI
AT A GLANCE
The report states that Aleph Alpha open-sourced the English-German bilingual MoE model Kolibri; each token activates approximately 3460000000 parameters.
Article
Europe’s open-source model Kolibri is here: 3.46B active parameters, supporting 1M context
Beating AI flash news, German AI company Aleph Alpha has released and open-sourced the English-German bilingual model Kolibri. The model uses MoE architecture, with a total of 78100000000, but each token activates approximately 3460000000 parameters. The weights have been uploaded to Hugging Face under the Apache 2.0 license.
Kolibri supports tool calling and adjustable reasoning modes, and its context can be extended to approximately 1000000 token s at most. However, the official recommendation is to keep everyday deployments within 262000 token s, and the longest context the model was actually trained on is also 262000.
In Aleph Alpha’s official evaluations, Kolibri achieved AIME 2025 and achieved 96.9%，LiveCodeBench v6. Its combined English-German score was slightly higher than that of 85.9%with approximately 3B active parameters, but it remained below the Qwen3.5 35B-A3B with 27B fully active, which achieved Qwen3.8.
Original link https://m.theblockbeats.info/flash/370241
Key points
01
Aleph Alpha has released and open-sourced the English-German bilingual model Kolibri.
02
Kolibri uses a MoE architecture, with a total of 78100000000, and each token activates approximately 3460000000 parameters.
03
The report states that the model weights have been uploaded to Hugging Face under the Apache 2.0 license.
04
Kolibri supports tool calling and adjustable reasoning modes.
05
Its context can be extended to approximately 1000000 token s at most, but the official recommendation is to keep it within 262000 token s for everyday deployments; the longest context actually used in training was also 262000。
06
Aleph Alpha’s official evaluation states that Kolibri achieved AIME 2025 and, on LiveCodeBench, 96.9%achieved v6. 85.9%。
07
The report states that Kolibri’s combined English-German score was slightly higher than that of Qwen3.5 35B-A3B, but lower than that of 27B with all parameters active, which achieved Qwen3.8。
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
MoE The architecture can be understood as the model having many parameters, while enabling only a portion of them when processing each token. Although Kolibri has a total of 78100000000 parameters, each token activates approximately 3460000000 parameters. The material also distinguishes between the maximum extendable context and the actual training length: the former is approximately 1000000 token s, while the latter is 262000 token s; the official recommendation for everyday deployment is also no more than 262000。
Why it matters to readers
This material presents an open-source model option supporting English-German bilingual use, tool calling, and long context;Apache 2.0 the Apache license and the uploaded weights are points of interest when assessing its usability.
Beginners should focus on distinguishing “total parameters” from “active parameters” and note that approximately 1000000 token s is the extendable limit, not the actual longest training length stated in the material. Benchmark scores come from official evaluations and cannot be directly equated with performance in all real-world use cases.
Risks and unknowns
The supplied material does not include a first-party announcement from Aleph Alpha or the Hugging Face model repository page; the weights, license, and specifications cannot be independently verified based on this evidence.
AIME 2025 and LiveCodeBench v6 scores are from Aleph Alpha’s official evaluation; the material does not provide independent retest results.
Related Developments
Loading event timeline…
The material does not explain the actual performance, resource requirements, or deployment costs under approximately 1000000 token s of context.
The material does not provide information on model adoption.
Related concepts
MoE
This term is not in the glossary yet. Browse related concepts in the glossary.
