FROMOpen weights, on their own metal

SAT · Oct 3, 2026Open weightsModels

Aleph Alpha released a 78B open-weight model built for sovereign deployment

Aleph Alpha released Kolibri on Saturday, the Day of German Reunification: an English-and-German mixture-of-experts model on Hugging Face under Apache 2.0. It carries 78.1 billion total parameters but activates about 3.46 billion per token, serves from roughly 78 GB of FP8 weights on a single H200 or B200, and holds a native 262,144-token context validated at 1,048,576. Training ran on 768 Nvidia B200s in Germany and Finland over 20 trillion tokens, roughly a fifth of them German.

Aleph Alpha’s own table puts Kolibri at 96.9 on AIME 2025 (87.5 German) and 84.3 on GPQA diamond, ahead of Qwen3.6-35B-A3B and Mistral Small 4 at similar active-parameter counts - while losing to Qwen3.6 on the τ²-bench telecom suite, 94.7 against 99.1, and on tool calling, 61.4 against 67.2. No independent ranking covers it yet, and all three models it is measured against date from the spring, so the open-weight leaders that shipped since go unmeasured. It came with a 189-page technical report, and the model card discloses the training text was rephrased with Google’s Gemma 4 and Mistral-NeMo, with Qwen3-32B labelling the quality filters.

Aleph Alpha’s Kolibri launch cover: the white Kolibri wordmark beside a white hummingbird icon on a green gradient, framed by thin stepped square outlines on the left and right
Aleph Alpha

WHY IT MATTERS

Sovereignty here is a procurement argument before it is a capability one: built and trained in Europe, no hosted API, and full weights a ministry or an aerospace supplier can run on its own metal and nobody can switch off. The engineering underneath is the active-parameter trick - compute like a 3.5-billion-parameter model while holding the memory of a 78-billion one - which buys benchmark parity with peers three to four times its active size at a serving cost a public administration can budget for. For anyone weighing open weights against a hosted API, the 189-page report is the more reproducible kind of release: the recipe, not just the weights. The caveat deserves one line: the scores are Aleph Alpha’s own, taken against peers from the spring rather than this month’s leaders, and serving past 262k tokens needs a flag.