Aleph Alpha Releases Kolibri-1: Open-Weights German AI Model with 78B Parameters



Aleph Alpha has released Kolibri-1, an open-weight AI model trained for German and English, with 78B total parameters and only 3.46B active per token. The company reports strong benchmark results, including 96.9% on AIME 2025.
Aleph Alpha Releases Kolibri-1
Aleph Alpha has introduced Kolibri-1, an open-weight AI model trained for German and English. The model has 78B total parameters, but only 3.46B parameters are activated for each token. The weights are available on Hugging Face: Kolibri-1 under the Apache 2.0 license. According to Aleph Alpha's own evaluations, Kolibri-1 scores 96.9% on AIME 2025 and 84.3% on GPQA Diamond. In overall German-language evaluation, it slightly outperforms Qwen3.5 35B-A3B and GPT-OSS 120B. The coding results are more mixed: 66.4% on SWE-Bench Verified compared with 73.8% for Qwen3.6. For post-training, the company says it used responses from Chinese models: GLM from Zhipu and Qwen from Alibaba, as described in the technical report. Kolibri-1 shows that competitive small models can come from outside the usual US and China ecosystems.
Why it matters
AnalysisKolibri-1 adds a European open-weight option for German and English workloads and demonstrates that smaller active-parameter models can reach strong benchmark results. It may matter for teams building localized AI products with Apache 2.0 licensing.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
Mistral AI (European Open and Efficient Models)
A series of high-performance open and commercial models from the French startup Mistral AI (Mistral 7B, Mixtral 8x7B, Mistral Large). Known for their compactness, speed, and respect for data privacy.
Open Weights vs Closed APIs
A comparison of two fundamental philosophies in AI distribution: Open Weights (Llama, Mistral, DeepSeek), which can be downloaded and run on your own server, versus Closed APIs (OpenAI, Anthropic, Google), where access is granted solely through subscription.
OpenAI Releases GPT 6.2 Astra and Sol
OpenAI has launched GPT 6.2 Astra and Sol, with Astra topping agent benchmarks and offering more competitive token pricing.
Part II. AI Agent Architecture: From Model to Memory and Tools
A deep engineering dive: tokenization, KV-cache, sampling determinism, the training stack, quantization, context engineering, the safety stack, production RAG, ReAct, and multi-agent swarm systems.