Aleph Alpha Drops Open 78B European Model

🔥 What happened Aleph Alpha released Kolibri, an open-weight German-English MoE model with 78.1B parameters that activates just 3.46B per token. Weights are on Hugging Face under Apache 2.0, trained on 768 B200 GPUs across nearly 24 trillion tokens. 💡 Why it matters For regulated industries, this means running on your own hardware instead of shipping data to a third party. The model abstains on 44% of unanswerable items versus 14.8% for its predecessor — a real grounding jump, plus 1M-token context and four reasoning levels. ⚡ Our take Finally a European model that doesn't just talk sovereignty but actually hands over the weights. Whether the vendor benchmarks survive independent testing is now the community's problem.